diff --git a/ACTIVE_SLICE.md b/ACTIVE_SLICE.md
index 59bc152..4d6b119 100644
--- a/ACTIVE_SLICE.md
+++ b/ACTIVE_SLICE.md
@@ -1,16 +1,13 @@
-# Active milestone
+# Milestone state
 
 ```text
-Milestone: 3 — independent projections
-Goal: Make manual output, portable graph artifacts, and the live viewer independent generation-pinned consumers of the validated graph.
-In scope: Versioned ManualRenderPlan and GraphViewPlan; immutable projection packages and receipts; independent manual and graph renderers; projection policies; incremental fragments; equivalence, recovery, accessibility, response-size, performance, and memory gates.
-Out of scope: Adapter SDK expansion; remote render services; shared render farms; third-party renderer ecosystems; self-hosting; storage replacement; WorldForge or ScrapeStation changes; production MCP repointing; tags and releases.
-Done when: Manual and graph plans are versioned and bounded; renderers cannot crawl project state or mutate canonical facts; portable and live graph modes remain separate; policies are enforced independently; status is receipt-only; full/incremental output is equivalent; accessibility and maintained scale gates pass.
-Status: Active implementation. Three independent audits were reconciled before source changes. The
-versioned plan/package/receipt contracts, pure manual and graph planners, isolated manual renderer,
-legacy byte-compatibility shim, packaged schemas, and pinned live-source correction are implemented
-and focused-green. Durable publication, portable graph artifacts, detached workers, fragment
-equivalence, independent policy enforcement, accessibility, and maintained scale gates remain.
+Last completed milestone: 3 — independent projections
+Outcome: Manual output, portable graph artifacts, and the live viewer are independent generation-pinned consumers of the validated graph.
+Evidence: Clean candidate f5dccb5e1c312121f1af63780162f593d9363b98; 281 tests and 272 subtests; 3 accessibility flows; clean 1,000-node ten-sample benchmark; isolated wheel proof; no secret-scan findings.
+Active milestone: None.
+Next directional milestone: 4 — adapter SDK and product documentation.
+Status: Milestone 3 is closed. Milestone 4 has not started.
 ```
 
-Milestones 4–5 remain directional context and are not active.
+Milestones 4–5 remain directional context. Do not begin Milestone 4 without a new active-slice
+contract.
diff --git a/DEVELOPMENT_NOTES.md b/DEVELOPMENT_NOTES.md
index f00abff..2548012 100644
--- a/DEVELOPMENT_NOTES.md
+++ b/DEVELOPMENT_NOTES.md
@@ -345,7 +345,7 @@ These are notes, not commitments:
   than bounded positions, they will need a different versioned security contract and persisted key
   lifecycle.
 
-## Milestone 2 — active: agent retrieval and MCP experience
+## Milestone 2 — complete: agent retrieval and MCP experience
 
 ### Audit reconciliation
 
@@ -611,10 +611,10 @@ task-shaped capsule for every continuation page, add authenticated continuation
 model requires it, verify a native Claude timeout representation, and introduce adapter-owned
 launcher metadata before generating configurations for custom adapters.
 
-## Milestone 3 — active: independent projections
+## Milestone 3 — complete: independent projections
 
 Milestone 3 began only after `main` and `dev` were aligned at the verified Milestone 2 closeout.
-Three read-only audits are running before source changes:
+Three read-only audits ran before source changes:
 
 - Manual planning, immutable packages, renderer isolation, receipts, preview/application
   integration, and full/incremental equivalence.
@@ -692,8 +692,95 @@ The first slice now implements:
 - A live-viewer correction: source evidence now comes from the pinned index generation. The
   viewer no longer reopens mutable canonical files behind an older graph snapshot.
 
-The new repository-native contract target passes 91 tests and 120 subtests. The combined
-projection, rendering, and live-viewer focus passes with byte-exact compatibility and no hidden
-source/path authority. This is not Milestone 3 closeout: durable multi-artifact publication,
-portable graph rendering, detached workers, fragment reuse/equivalence, policy version 2,
-accessibility, and maintained scale evidence remain active work.
+The new repository-native contract target passed 91 tests and 120 subtests at the slice boundary.
+The combined projection, rendering, and live-viewer focus passed with byte-exact compatibility and
+no hidden source/path authority.
+
+### Durable portable graph publication
+
+The portable graph path now has its own declared `graph_render` views, pure plans, fixed
+`portable_graph_html` renderer, content-addressed artifact store, renderer receipts, and one bounded
+generation/view manifest as the publication commit. It supports Nodes, Flow, and Web without
+including Logic. Static HTML contains the complete pre-rendered graph and treats JavaScript as
+progressive enhancement.
+
+Publication revalidates source, view, artifact, receipt, and output identities across replacement.
+Status reads only bounded manifest and receipt evidence. It never plans or renders. Repair may
+restore a declared output from its content-addressed artifact. A post-artifact failure that cannot
+be rolled back returns explicit degraded committed evidence rather than reporting an ordinary
+failed mutation.
+
+### Detached workers and incremental fragments
+
+Manual and portable graph packages execute through one fixed one-request child protocol. The
+parent launches isolated Python from a trusted working directory with a sanitized environment,
+spools stdout to disk, reads one bounded canonical response, and validates the complete artifact
+and receipt identity. The worker accepts only the two built-in renderer identities. Requests are
+bounded by the 24,000,000-byte package contract, actual artifact transfer by 20,000,000 bytes, and
+execution by a 30-second timeout.
+
+Manual fragment records are semantic, versioned, canonical, hash-bound, and stored below a
+dedicated confined cache. The worker independently recomputes the expected page fragment before
+using a record. Corrupt, forged, oversized, stale, or aggregate-oversized records fall back to the
+full detached render. Cold fragment creation is compared byte-for-byte with that full oracle before
+cache publication. The cache retains only the current inventory and is capped at 10,000 entries
+and 64,000,000 bytes.
+
+### Independent policies and accessibility
+
+Projection policy version 2 independently composes manual `auto|explicit|disabled`, portable graph
+`explicit|disabled`, and live viewer `on-demand|disabled`. CLI, MCP, generated client
+configuration, doctor, render services, canonical application, onboarding, and viewer-manager
+entry points enforce their relevant policy. Status remains available when an active operation is
+disabled.
+
+Generated client evidence binds the projection policy, its hash, projection availability, and the
+current descriptor hash into the configuration hash. Validation cross-checks omitted default
+selectors against the bound descriptor so coordinated policy and availability drift fails closed.
+The version-1 effective-policy payload remains unchanged for existing clients.
+
+Pinned Playwright 1.62.0 and axe-core 4.12.1 gates exercise the frozen manual, portable graph, and
+live viewer with selected WCAG A/AA axe tags and keyboard interaction flows. Portable and live
+graph presentation received only the minimal contrast and nested-role corrections needed by those
+gates.
+
+### Scale and runtime hardening
+
+The first 1,000-node full benchmark exposed recursive strongly connected-component traversal in
+manual planning. Cycle detection now uses an iterative two-pass traversal. A regression covers the
+descriptor maximum of 10,000 nodes as both a deep acyclic chain and one strongly connected
+component.
+
+The isolated wheel proof also exposed a Python `runpy` warning when the worker module was imported
+during package initialization before `-m` execution. A private fixed module entrypoint now owns
+child startup. Malformed child input returns code 2 with empty stdout and stderr.
+
+Configured render ceilings above 20,000,000 bytes remain accepted for compatibility, and small
+actual artifacts render normally. The detached protocol still rejects an actual transfer beyond
+its fixed 20,000,000-byte boundary.
+
+### Milestone 3 closeout
+
+Candidate `f5dccb5e1c312121f1af63780162f593d9363b98` passed the complete repository gate: formatting,
+Python and web lint, strict types, compilation, 281 tests and 272 subtests, three accessibility
+flows, lock and dependency checks, package builds, and all milestone smoke benchmarks. The
+maintained projection contract subset passed 142 tests and 236 subtests.
+
+The clean ten-sample 1,000-node benchmark passed every latency, memory, response-size, no-work, and
+equivalence gate. Manual full rendering measured 810.490 ms p95, portable graph full rendering
+323.690 ms p95, and receipt-only status 111.381 ms and 59.331 ms p95 respectively. Direct detached
+worker peaks were 88,580,096 and 89,583,616 bytes. The separately gated production manual worker
+peak was 104,771,584 bytes. Production cold, warm, forced-full, add, change, delete, and reorder
+outputs were byte-identical.
+
+Production warm fragment rendering measured 2,206.540 ms p95 versus 978.870 ms for forced full.
+Milestone 3 therefore closes the fragment isolation, invalidation, equivalence, and recovery
+contract without claiming a throughput win. Later optimization must begin from that evidence.
+
+The exact method and measurements are recorded in `docs/MILESTONE_3_BASELINE.md` and
+`benchmarks/milestone3-2026-07-29.json`. The candidate passed an isolated wheel CLI/MCP/worker
+proof. Gitleaks 8.30.1 found no findings across the six Milestone 3 commits or candidate tree.
+
+Milestone 3 is complete. No tag, release, production integration repointing, WorldForge change,
+ScrapeStation change, storage rewrite, or self-hosting dependency was introduced. Milestone 4
+remains directional and has not started.
diff --git a/README.md b/README.md
index 16a03a6..20ba862 100644
--- a/README.md
+++ b/README.md
@@ -26,6 +26,12 @@ declared manuals, visualizes project structure, and manages reviewable documenta
 - Detects project-local adapter implementation and configuration changes and requires a fresh
   project-bound process before any further MCP work.
 - Keeps function-scoped control-flow projections separate from the primary architecture graph.
+- Compiles manuals and portable graph artifacts from separate versioned, generation-pinned plans
+  and immutable packages.
+- Runs built-in manual and graph renderers in fixed detached workers with validated receipts,
+  bounded transfer, and no project-path authority.
+- Enforces independent manual, portable-graph, and live-viewer policy while keeping status
+  receipt-only.
 - Runs a managed loopback graph browser with neighborhood, semantic Flow, convergence Web,
   function-scoped Logic, source inspection, and branch-aware node hiding.
 - Supports generic documentation projects and project-owned source adapters.
@@ -73,6 +79,63 @@ details, and explicit truncation. Paged results use one top-level cursor and a v
 `receipt_header`; `stored_receipt_hash` identifies the complete persisted receipt. The read never
 exposes Logic details, loads canonical source, repairs derived state, or invents history.
 
+## Independent projections
+
+Manual compilation, portable graph rendering, and the live viewer consume the same validated graph
+generation through separate boundaries:
+
+```text
+validated generation
+    ├── ManualRenderPlanV1 → immutable package → detached manual renderer
+    ├── GraphViewPlanV1    → immutable package → detached portable graph renderer
+    └── pinned index       → managed read-only live viewer
+```
+
+Plans, packages, and receipts are canonical, versioned, hash-identified, bounded, and contain no
+project object, SQLite handle, absolute project path, command, or caller-selected renderer module.
+Renderers cannot select graph facts, crawl canonical sources, choose publication paths, or mutate
+project state.
+
+Declare portable graph output separately from manual views:
+
+```toml
+[graph_render]
+output_root = ".docforge/portable-graph"
+
+[[graph_render.views]]
+id = "architecture"
+renderer = "portable_graph_html"
+output = "architecture.html"
+title = "Architecture"
+root = "architecture.overview"
+initial_mode = "web"
+depth = 3
+max_nodes = 250
+max_edges = 1000
+max_work = 100000
+include_logic = false
+```
+
+Plan, publish, and inspect it explicitly:
+
+```bash
+.venv/bin/docforge --project-root "$PROJECT" graph-plan architecture
+.venv/bin/docforge --project-root "$PROJECT" graph-render architecture
+.venv/bin/docforge --project-root "$PROJECT" graph-render-status architecture
+```
+
+The version-2 projection policy independently selects manual
+`auto|explicit|disabled`, portable graph `explicit|disabled`, and live viewer
+`on-demand|disabled`. Use `--manual-render-policy`, `--portable-graph-policy`, and
+`--live-viewer-policy` on CLI/MCP startup or generated client configuration. Status remains
+available when the corresponding active operation is disabled.
+
+Manual fragment reuse is disposable. Cold record publication is guarded by byte-exact comparison
+with a full detached render; warm records are independently recomputed and validated inside the
+worker. Full rendering remains the recovery and equivalence oracle. Portable artifacts commit
+content-addressed output and renderer evidence before one bounded generation/view manifest; status
+does not plan or render.
+
 ## Graph views
 
 The browser presents the primary architecture graph through three complementary views and loads a
@@ -170,10 +233,18 @@ DocForge describes them as a source graph.
   performance, memory, rendering and response sizes, bottlenecks, and missing coverage.
 - [Milestone 0 closeout](docs/MILESTONE_0_CLOSEOUT.md) — lineage, migration, security scan,
   repository state, and fresh-clone proof.
+- [Milestone 1 baseline](docs/MILESTONE_1_BASELINE.md) — warm operation latency, structured work,
+  status, retrieval, and memory measurements.
+- [Milestone 1 closeout](docs/MILESTONE_1_CLOSEOUT.md) — fast-core contracts, adversarial
+  validation, compatibility boundaries, and exact candidate evidence.
 - [Milestone 2 baseline](docs/MILESTONE_2_BASELINE.md) — task context, generation diff, client
   configuration, doctor, response-size, counter, and memory measurements.
 - [Milestone 2 closeout](docs/MILESTONE_2_CLOSEOUT.md) — implemented contracts, adversarial
   validation, exclusions, and exact candidate evidence.
+- [Milestone 3 baseline](docs/MILESTONE_3_BASELINE.md) — manual, fragment, worker, portable graph,
+  status, equivalence, response-size, and memory measurements.
+- [Milestone 3 closeout](docs/MILESTONE_3_CLOSEOUT.md) — independent projection contracts,
+  adversarial validation, compatibility boundaries, and exact candidate evidence.
 - [MCP contract](docs/MCP_CONTRACT.md) — exact tool and process boundary.
 - [Viewer manager](docs/VIEWER_MANAGER.md) — native service setup and lifecycle.
 - [Adapter decision](docs/APPLICATION_DECISION.md) — why custom adapters own canonical
@@ -197,6 +268,8 @@ make gate
 Focused entry points are available as `make contract`, `make test`, `make type`,
 `make benchmark-smoke`, `make benchmark`, `make benchmark-m1-smoke`, and
 `make benchmark-m1`. Milestone 2 adds `make benchmark-m2-smoke` and `make benchmark-m2`.
+Milestone 3 adds `make accessibility`, `make benchmark-m3-smoke`, `make benchmark-m3`, and
+`make benchmark-m3-full`.
 
 The committed 1,000-node baseline and its measurement method are under `benchmarks/`.
 
diff --git a/SLICE_HISTORY.md b/SLICE_HISTORY.md
index fa7bf77..5b4cd75 100644
--- a/SLICE_HISTORY.md
+++ b/SLICE_HISTORY.md
@@ -1,5 +1,51 @@
 # Completed slices
 
+## DocForge2 Milestone 3 independent projections
+
+### Changed
+
+- Added strict version-1 manual plan, graph plan, projection package, and projection receipt
+  contracts with canonical identities and packaged schemas.
+- Kept the legacy manual API and exact alpha bytes as a compatibility wrapper over a pure planner,
+  immutable package, and independent renderer.
+- Added generation-pinned portable graph planning, detached rendering, content-addressed durable
+  publication, repair, and receipt-only status.
+- Added a fixed isolated worker protocol with bounded request, response, artifact, timeout,
+  environment, and renderer inventory.
+- Added semantic manual fragments with independent worker validation, corruption recovery,
+  full-render equivalence, and bounded cache retention.
+- Added version-2 independent manual, portable-graph, and live-viewer policy with descriptor-bound
+  generated-client evidence while retaining effective policy version 1.
+- Corrected live source reads to use the pinned index generation.
+- Added static and interactive accessibility gates for the manual, portable graph, and live viewer.
+- Replaced recursive cycle planning with an iterative traversal proven at 10,000 nodes.
+
+### Verification
+
+- The complete gate passed 281 tests, 272 subtests, three accessibility flows, dependency checks,
+  builds, and all milestone smoke benchmarks.
+- The maintained projection contract subset passed 142 tests and 236 subtests.
+- The clean ten-sample 1,000-node benchmark passed every latency, memory, response-size, no-work,
+  deterministic-output, and full/incremental equivalence threshold.
+- An isolated wheel passed CLI, MCP, detached manual rendering, and malformed worker startup.
+- Three adversarial review tracks were reconciled, and Gitleaks found no findings in the six
+  milestone commits or candidate tree.
+
+### Limits
+
+- Production fragment reuse is correct and recoverable but is slower than forced full rendering at
+  the maintained 1,000-page fixture. No speedup is claimed.
+- Actual detached artifact transfer is capped at 20,000,000 bytes.
+- Portable graph publication is CLI-owned. MCP exposes read-only plan and status tools.
+- Remote render services, render farms, third-party renderer ecosystems, storage replacement,
+  self-hosting, adapter SDK expansion, and production integration changes remain out of scope.
+- No tag or release was created.
+
+### Next gate
+
+Milestone 4 remains directional and is not active. Create a new active-slice contract before
+starting adapter SDK or product-documentation implementation.
+
 ## DocForge2 Milestone 0 successor foundation
 
 ### Changed
diff --git a/benchmarks/README.md b/benchmarks/README.md
index 9d78eda..2141f1a 100644
--- a/benchmarks/README.md
+++ b/benchmarks/README.md
@@ -34,6 +34,14 @@ make benchmark-m2-smoke
 make benchmark-m2
 ```
 
+Run the Milestone 3 independent-projection smoke and maintained full gates:
+
+```bash
+make benchmark-m3-smoke
+make benchmark-m3
+make benchmark-m3-full
+```
+
 The benchmark creates canonical sources, derived state, changesets, rendered output, and caches
 only in a disposable temporary directory. It does not read another project, self-host DocForge, or
 mutate repository content.
@@ -56,9 +64,16 @@ records whether diagnostics were dropped for response budget, checks all hidden-
 and measures isolated-process peak RSS. Its interpretation is in
 [`docs/MILESTONE_2_BASELINE.md`](../docs/MILESTONE_2_BASELINE.md).
 
-The generic fixture exposes whole-source scaling. It does not replace the incremental adapter
-equivalence tests and does not claim to measure a portable graph renderer, because Milestone 0 has
-no portable graph-planning or graph-rendering contract.
+`milestone3-2026-07-29.json` is the clean-tree independent-projection baseline captured from commit
+`f5dccb5e1c312121f1af63780162f593d9363b98`. It measures versioned manual and graph planning,
+in-process and detached rendering, production cold/warm/forced-full behavior, fragment cache
+sweeps, add/change/delete/reorder equivalence, portable publication, receipt-only status, traced
+memory, detached worker peak memory, and response size. Its interpretation is in
+[`docs/MILESTONE_3_BASELINE.md`](../docs/MILESTONE_3_BASELINE.md).
+
+The generic fixtures expose whole-source and projection scaling. They do not replace incremental
+adapter equivalence tests. Milestone 3's portable fixture contains 1,000 Nodes/Flow/Web nodes and
+999 edges; portable version 1 deliberately excludes Logic.
 
 The Milestone 1 harness treats wall time and structured work counters as separate gates. Warm
 operations fail if they load a complete project, parse source files, reconstruct an adapter
@@ -73,4 +88,5 @@ status. The 1,000-node run records bounded semantic summaries for exact errors,
 backlinks, both traversal directions, paged context, render receipt states, and visualization
 freshness. The reported p95 uses the nearest-rank method; with ten samples it is the maximum.
 `process_peak_rss_kib` is the cumulative main-process `RUSAGE_SELF` high-water mark and excludes the
-detached viewer worker.
+detached viewer worker. It is diagnostic and not operation-local. Milestone 3 memory gates use
+per-operation `tracemalloc` peaks and detached worker receipt peaks instead.
diff --git a/benchmarks/milestone3-2026-07-29.json b/benchmarks/milestone3-2026-07-29.json
new file mode 100644
index 0000000..3386a8d
--- /dev/null
+++ b/benchmarks/milestone3-2026-07-29.json
@@ -0,0 +1,442 @@
+{
+  "benchmark": "docforge2_milestone3",
+  "environment": {
+    "implementation": "CPython",
+    "machine": "x86_64",
+    "platform": "Linux-7.1.3-200.nobara.fc44.x86_64-x86_64-with-glibc2.43",
+    "python": "3.14.6"
+  },
+  "equivalence": {
+    "manual_full_vs_fragment_assisted": true,
+    "manual_in_process_vs_detached": true,
+    "manual_production_cold_warm_full": true,
+    "manual_production_variants": {
+      "add": true,
+      "change": true,
+      "delete": true,
+      "reorder": true
+    },
+    "portable_graph_in_process_vs_detached": true
+  },
+  "fixture": {
+    "edge_count": 999,
+    "full_coverage": true,
+    "kind": "synthetic_generic_projection",
+    "manual_page_count": 1000,
+    "node_count": 1000,
+    "portable_graph_edge_count": 999,
+    "portable_graph_node_count": 1000
+  },
+  "memory": {
+    "manual_production_worker_peak_bytes": 104771584,
+    "manual_worker_peak_bytes": 88580096,
+    "portable_graph_worker_peak_bytes": 89583616,
+    "process_peak_rss_kib": 102692
+  },
+  "method": {
+    "clock": "time.perf_counter_ns",
+    "detached_peak_memory": "worker receipt resource peak RSS",
+    "determinism": "stable semantic summaries must match across samples; report JSON uses sorted keys",
+    "full_mode_node_requirement": 1000,
+    "in_process_peak_memory": "tracemalloc per measured invocation",
+    "maximum_worker_artifact_bytes": 20000000,
+    "process_peak_memory": "resource.getrusage(RUSAGE_SELF).ru_maxrss",
+    "response_size": "UTF-8 bytes of canonical compact sorted JSON",
+    "samples": 10
+  },
+  "mode": "full",
+  "operations": {
+    "fragment_assisted_equivalence": {
+      "max_ms": 666.119,
+      "maximum_response_bytes": 664,
+      "maximum_traced_peak_bytes": 5332792,
+      "median_ms": 638.999,
+      "min_ms": 633.157,
+      "p95_limit_ms": 15000,
+      "p95_ms": 666.119,
+      "response_limit_bytes": 128000,
+      "samples": 10,
+      "stable_result": {
+        "artifact_bytes": 583149,
+        "artifact_sha256": "278ff8cd6b092e7d8114527dba88051917aa827b22d4aed61e57acf004e5870c",
+        "package_bytes": 2364302,
+        "package_id": "ecc21e58106c420a9f78ffbba997778ca98b61586ecfdb23a15b08f431a52c42",
+        "plan_bytes": 1006393,
+        "plan_id": "47008f5177230e3ad5abbfee8bccaff0498963106cb33ed7cdd6f8bbc3c495de"
+      },
+      "traced_peak_limit_bytes": 268435456
+    },
+    "fragment_cache_hit_sweep": {
+      "max_ms": 514.324,
+      "maximum_response_bytes": 145,
+      "maximum_traced_peak_bytes": 6941174,
+      "median_ms": 498.911,
+      "min_ms": 495.447,
+      "p95_limit_ms": 5000,
+      "p95_ms": 514.324,
+      "response_limit_bytes": 32768,
+      "samples": 10,
+      "stable_result": {
+        "aggregate_content_bytes": 516921,
+        "fragment_count": 1000,
+        "ordered_record_hash": "99a95d575504238233a6367c884fa1d50231855f2f2ffe3dae90fa18c4136f72"
+      },
+      "traced_peak_limit_bytes": 268435456
+    },
+    "fragment_cache_miss_sweep": {
+      "max_ms": 106.888,
+      "maximum_response_bytes": 145,
+      "maximum_traced_peak_bytes": 146245,
+      "median_ms": 106.203,
+      "min_ms": 105.383,
+      "p95_limit_ms": 5000,
+      "p95_ms": 106.888,
+      "response_limit_bytes": 32768,
+      "samples": 10,
+      "stable_result": {
+        "aggregate_content_bytes": 516921,
+        "fragment_count": 1000,
+        "ordered_record_hash": "99a95d575504238233a6367c884fa1d50231855f2f2ffe3dae90fa18c4136f72"
+      },
+      "traced_peak_limit_bytes": 268435456
+    },
+    "fragment_cache_put_sweep": {
+      "max_ms": 539.212,
+      "maximum_response_bytes": 145,
+      "maximum_traced_peak_bytes": 509548,
+      "median_ms": 539.212,
+      "min_ms": 539.212,
+      "p95_limit_ms": 10000,
+      "p95_ms": 539.212,
+      "response_limit_bytes": 32768,
+      "samples": 1,
+      "stable_result": {
+        "aggregate_content_bytes": 516921,
+        "fragment_count": 1000,
+        "ordered_record_hash": "99a95d575504238233a6367c884fa1d50231855f2f2ffe3dae90fa18c4136f72"
+      },
+      "traced_peak_limit_bytes": 268435456
+    },
+    "manual_detached_worker": {
+      "child_peak_limit_bytes": 268435456,
+      "max_ms": 599.394,
+      "maximum_child_peak_bytes": 88580096,
+      "maximum_response_bytes": 667,
+      "maximum_traced_peak_bytes": 28203708,
+      "median_ms": 591.317,
+      "min_ms": 585.123,
+      "p95_limit_ms": 20000,
+      "p95_ms": 599.394,
+      "response_limit_bytes": 128000,
+      "samples": 10,
+      "stable_result": {
+        "artifact_bytes": 583149,
+        "artifact_sha256": "278ff8cd6b092e7d8114527dba88051917aa827b22d4aed61e57acf004e5870c",
+        "package_bytes": 1007297,
+        "package_id": "5fec20ace40b6db7e89083704af371245a5b5b73247f71edb6eab0cb2bd5e2c1",
+        "plan_bytes": 1006393,
+        "plan_id": "47008f5177230e3ad5abbfee8bccaff0498963106cb33ed7cdd6f8bbc3c495de"
+      },
+      "traced_peak_limit_bytes": 268435456
+    },
+    "manual_forced_full": {
+      "child_peak_limit_bytes": 268435456,
+      "max_ms": 978.87,
+      "maximum_child_peak_bytes": 104771584,
+      "maximum_response_bytes": 668,
+      "maximum_traced_peak_bytes": 30425714,
+      "median_ms": 944.135,
+      "min_ms": 934.773,
+      "p95_limit_ms": 20000,
+      "p95_ms": 978.87,
+      "response_limit_bytes": 128000,
+      "samples": 10,
+      "stable_result": {
+        "output_bytes": 583149,
+        "output_sha256": "278ff8cd6b092e7d8114527dba88051917aa827b22d4aed61e57acf004e5870c",
+        "render_identity": "73f9a175afd7bfea5af70d4c7df902c982f042c413d533817dc2bade9e54d4b0"
+      },
+      "traced_peak_limit_bytes": 268435456
+    },
+    "manual_full_render": {
+      "max_ms": 810.49,
+      "maximum_response_bytes": 664,
+      "maximum_traced_peak_bytes": 4412754,
+      "median_ms": 801.948,
+      "min_ms": 773.7,
+      "p95_limit_ms": 15000,
+      "p95_ms": 810.49,
+      "response_limit_bytes": 128000,
+      "samples": 10,
+      "stable_result": {
+        "artifact_bytes": 583149,
+        "artifact_sha256": "278ff8cd6b092e7d8114527dba88051917aa827b22d4aed61e57acf004e5870c",
+        "package_bytes": 1007297,
+        "package_id": "5fec20ace40b6db7e89083704af371245a5b5b73247f71edb6eab0cb2bd5e2c1",
+        "plan_bytes": 1006393,
+        "plan_id": "47008f5177230e3ad5abbfee8bccaff0498963106cb33ed7cdd6f8bbc3c495de"
+      },
+      "traced_peak_limit_bytes": 268435456
+    },
+    "manual_incremental_cold": {
+      "child_peak_limit_bytes": 268435456,
+      "max_ms": 2827.152,
+      "maximum_child_peak_bytes": 99454976,
+      "maximum_response_bytes": 668,
+      "maximum_traced_peak_bytes": 35160716,
+      "median_ms": 2827.152,
+      "min_ms": 2827.152,
+      "p95_limit_ms": 20000,
+      "p95_ms": 2827.152,
+      "response_limit_bytes": 128000,
+      "samples": 1,
+      "stable_result": {
+        "output_bytes": 583149,
+        "output_sha256": "278ff8cd6b092e7d8114527dba88051917aa827b22d4aed61e57acf004e5870c",
+        "render_identity": "73f9a175afd7bfea5af70d4c7df902c982f042c413d533817dc2bade9e54d4b0"
+      },
+      "traced_peak_limit_bytes": 268435456
+    },
+    "manual_incremental_warm": {
+      "child_peak_limit_bytes": 268435456,
+      "max_ms": 2206.54,
+      "maximum_child_peak_bytes": 104767488,
+      "maximum_response_bytes": 669,
+      "maximum_traced_peak_bytes": 34676043,
+      "median_ms": 2143.388,
+      "min_ms": 2125.466,
+      "p95_limit_ms": 20000,
+      "p95_ms": 2206.54,
+      "response_limit_bytes": 128000,
+      "samples": 10,
+      "stable_result": {
+        "output_bytes": 583149,
+        "output_sha256": "278ff8cd6b092e7d8114527dba88051917aa827b22d4aed61e57acf004e5870c",
+        "render_identity": "73f9a175afd7bfea5af70d4c7df902c982f042c413d533817dc2bade9e54d4b0"
+      },
+      "traced_peak_limit_bytes": 268435456
+    },
+    "manual_status_no_work": {
+      "max_ms": 111.381,
+      "maximum_response_bytes": 1542,
+      "maximum_traced_peak_bytes": 1792025,
+      "median_ms": 62.881,
+      "min_ms": 60.111,
+      "p95_limit_ms": 500,
+      "p95_ms": 111.381,
+      "response_limit_bytes": 256000,
+      "samples": 10,
+      "stable_result": {
+        "counters": {
+          "adapter_projection_loads": 0,
+          "adapter_source_extractions": 0,
+          "index_builds": 0,
+          "index_checks": 0,
+          "index_synchronizations": 0,
+          "project_loads": 0,
+          "render_output_bytes_built": 0,
+          "render_output_bytes_hashed": 0,
+          "render_prepare_calls": 0,
+          "source_bytes_parsed": 0,
+          "source_files_parsed": 0,
+          "source_generation_checks": 2,
+          "viewer_manager_requests": 0
+        },
+        "response": {
+          "adapter": "generic",
+          "configured": true,
+          "outputs": [
+            {
+              "actual_output_hash": "278ff8cd6b092e7d8114527dba88051917aa827b22d4aed61e57acf004e5870c",
+              "expected_output_hash": "278ff8cd6b092e7d8114527dba88051917aa827b22d4aed61e57acf004e5870c",
+              "path": ".docforge/rendered/manual.html",
+              "projection_receipt": {
+                "artifacts": [
+                  {
+                    "artifact_id": "manual.html",
+                    "bytes": 583149,
+                    "media_type": "text/html; charset=utf-8",
+                    "sha256": "278ff8cd6b092e7d8114527dba88051917aa827b22d4aed61e57acf004e5870c"
+                  }
+                ],
+                "contract": "docforge.projection-receipt",
+                "diagnostics": {
+                  "warnings": []
+                },
+                "kind": "manual",
+                "package_id": "b30c65390570537a16ce03afe6a992599c6ca0c4ef5ea2f06d14b33e248bc9aa",
+                "peak_memory_bytes": 104771584,
+                "plan_id": "47008f5177230e3ad5abbfee8bccaff0498963106cb33ed7cdd6f8bbc3c495de",
+                "receipt_id": "ead4c25b34ada8df674903a592e25dd6bbf9817df2b04e5ebc66cde9e41c1ed1",
+                "renderer": {
+                  "renderer_id": "generic_html",
+                  "renderer_version": "1+markdown-it-py-4.2.0"
+                },
+                "schema_version": 1,
+                "timing": {
+                  "elapsed_ns": 114544865
+                }
+              },
+              "reason": null,
+              "receipt_schema_version": 1,
+              "render_identity": "73f9a175afd7bfea5af70d4c7df902c982f042c413d533817dc2bade9e54d4b0",
+              "renderer": "generic_html",
+              "renderer_version": "1+markdown-it-py-4.2.0",
+              "state": "current",
+              "template_hash": "ae0ebb3eadeb530e9d033c8fbd21321d406e29d04c0ae7d07a9a0f929f0ffa9f",
+              "verification": "receipt",
+              "view_id": "manual"
+            }
+          ],
+          "project_id": "synthetic-1000",
+          "project_root_fingerprint": "0747acfd975703c0",
+          "revision": "unversioned",
+          "source_hash": "a314da4ffa8fcf291ef7a7b0fc87737caeaa89fa3c558ea442afeb0e5ae49c2d",
+          "state": "current",
+          "status": "ok",
+          "verification": "receipt"
+        }
+      },
+      "traced_peak_limit_bytes": 268435456
+    },
+    "portable_graph_detached_worker": {
+      "child_peak_limit_bytes": 268435456,
+      "max_ms": 268.428,
+      "maximum_child_peak_bytes": 89583616,
+      "maximum_response_bytes": 660,
+      "maximum_traced_peak_bytes": 27587227,
+      "median_ms": 265.448,
+      "min_ms": 263.761,
+      "p95_limit_ms": 20000,
+      "p95_ms": 268.428,
+      "response_limit_bytes": 128000,
+      "samples": 10,
+      "stable_result": {
+        "artifact_bytes": 718383,
+        "artifact_sha256": "91dd5bb31679225e5af40a1f94a9e75f72921e36416e3d8b00c45f46a37a3a30",
+        "package_bytes": 398715,
+        "package_id": "72f96d7ab92c1d6b59f4f32539558232ecf35deda099fc4887c77890da1d4309",
+        "plan_bytes": 398158,
+        "plan_id": "6dac35937d855f72c6670f87aace3c0cb1dbff67bb842c55370429584064f76b"
+      },
+      "traced_peak_limit_bytes": 268435456
+    },
+    "portable_graph_full_render": {
+      "max_ms": 323.69,
+      "maximum_response_bytes": 657,
+      "maximum_traced_peak_bytes": 3490976,
+      "median_ms": 303.736,
+      "min_ms": 301.215,
+      "p95_limit_ms": 10000,
+      "p95_ms": 323.69,
+      "response_limit_bytes": 128000,
+      "samples": 10,
+      "stable_result": {
+        "artifact_bytes": 718383,
+        "artifact_sha256": "91dd5bb31679225e5af40a1f94a9e75f72921e36416e3d8b00c45f46a37a3a30",
+        "package_bytes": 398715,
+        "package_id": "72f96d7ab92c1d6b59f4f32539558232ecf35deda099fc4887c77890da1d4309",
+        "plan_bytes": 398158,
+        "plan_id": "6dac35937d855f72c6670f87aace3c0cb1dbff67bb842c55370429584064f76b"
+      },
+      "traced_peak_limit_bytes": 268435456
+    },
+    "portable_graph_status_no_work": {
+      "max_ms": 59.331,
+      "maximum_response_bytes": 879,
+      "maximum_traced_peak_bytes": 685012,
+      "median_ms": 58.467,
+      "min_ms": 58.052,
+      "p95_limit_ms": 500,
+      "p95_ms": 59.331,
+      "response_limit_bytes": 256000,
+      "samples": 10,
+      "stable_result": {
+        "counters": {
+          "adapter_projection_loads": 0,
+          "adapter_source_extractions": 0,
+          "index_builds": 0,
+          "index_checks": 0,
+          "index_synchronizations": 0,
+          "project_loads": 0,
+          "render_output_bytes_built": 0,
+          "render_output_bytes_hashed": 0,
+          "render_prepare_calls": 0,
+          "source_bytes_parsed": 0,
+          "source_files_parsed": 0,
+          "source_generation_checks": 2,
+          "viewer_manager_requests": 0
+        },
+        "response": {
+          "adapter": "generic",
+          "configured": true,
+          "outputs": [
+            {
+              "artifact": {
+                "artifact_id": "portable-graph.html",
+                "bytes": 718383,
+                "media_type": "text/html; charset=utf-8",
+                "sha256": "91dd5bb31679225e5af40a1f94a9e75f72921e36416e3d8b00c45f46a37a3a30"
+              },
+              "package_id": "72f96d7ab92c1d6b59f4f32539558232ecf35deda099fc4887c77890da1d4309",
+              "path": ".docforge/portable-graph/architecture.html",
+              "plan_id": "6dac35937d855f72c6670f87aace3c0cb1dbff67bb842c55370429584064f76b",
+              "publication_id": "2620ce106593d5499931b145a42fc1eccd31f5e356b19bf1d76687f8d45bcad2",
+              "reason": null,
+              "renderer": "portable_graph_html",
+              "renderer_version": "1",
+              "state": "current",
+              "verification": "manifest",
+              "view_id": "architecture"
+            }
+          ],
+          "project_id": "synthetic-1000",
+          "project_root_fingerprint": "0747acfd975703c0",
+          "revision": "unversioned",
+          "source_hash": "a314da4ffa8fcf291ef7a7b0fc87737caeaa89fa3c558ea442afeb0e5ae49c2d",
+          "state": "current",
+          "status": "ok"
+        }
+      },
+      "traced_peak_limit_bytes": 268435456
+    }
+  },
+  "schema_version": 1,
+  "sizes": {
+    "fragment_assisted_manual": {
+      "aggregate_fragment_content_bytes": 516921,
+      "artifact_bytes": 583149,
+      "artifact_sha256": "278ff8cd6b092e7d8114527dba88051917aa827b22d4aed61e57acf004e5870c",
+      "fragment_count": 1000,
+      "package_bytes": 2364302,
+      "package_id": "ecc21e58106c420a9f78ffbba997778ca98b61586ecfdb23a15b08f431a52c42",
+      "plan_bytes": 1006393,
+      "plan_id": "47008f5177230e3ad5abbfee8bccaff0498963106cb33ed7cdd6f8bbc3c495de",
+      "receipt_bytes": 664
+    },
+    "manual": {
+      "artifact_bytes": 583149,
+      "artifact_sha256": "278ff8cd6b092e7d8114527dba88051917aa827b22d4aed61e57acf004e5870c",
+      "package_bytes": 1007297,
+      "package_id": "5fec20ace40b6db7e89083704af371245a5b5b73247f71edb6eab0cb2bd5e2c1",
+      "plan_bytes": 1006393,
+      "plan_id": "47008f5177230e3ad5abbfee8bccaff0498963106cb33ed7cdd6f8bbc3c495de",
+      "receipt_bytes": 664
+    },
+    "manual_status_response_bytes": 1542,
+    "portable_graph": {
+      "artifact_bytes": 718383,
+      "artifact_sha256": "91dd5bb31679225e5af40a1f94a9e75f72921e36416e3d8b00c45f46a37a3a30",
+      "package_bytes": 398715,
+      "package_id": "72f96d7ab92c1d6b59f4f32539558232ecf35deda099fc4887c77890da1d4309",
+      "plan_bytes": 398158,
+      "plan_id": "6dac35937d855f72c6670f87aace3c0cb1dbff67bb842c55370429584064f76b",
+      "receipt_bytes": 657
+    },
+    "portable_graph_status_response_bytes": 879
+  },
+  "source": {
+    "dirty": false,
+    "revision": "f5dccb5e1c312121f1af63780162f593d9363b98"
+  }
+}
diff --git a/docs/COMPATIBILITY.md b/docs/COMPATIBILITY.md
index 12c6247..3925f59 100644
--- a/docs/COMPATIBILITY.md
+++ b/docs/COMPATIBILITY.md
@@ -71,6 +71,19 @@ Existing hand-written client configurations remain valid and are never rewritten
 Doctor is inspection-only and does not become a hidden bootstrap, synchronization, or migration
 path.
 
+The following Milestone 3 CLI additions are also additive:
+
+- `docforge graph-plan VIEW_ID`
+- `docforge graph-render VIEW_ID`
+- `docforge graph-render-status [VIEW_ID]`
+- `--manual-render-policy auto|explicit|disabled`
+- `--portable-graph-policy explicit|disabled`
+- `--live-viewer-policy on-demand|disabled`
+
+MCP adds the read-only `docforge_graph_plan` and `docforge_graph_render_status` tools. Portable
+graph publication remains an explicit local CLI integration action. Existing manual render,
+preview, visualization, and status names remain supported.
+
 ## Versioned data contracts
 
 Milestone 0 preserves:
@@ -96,6 +109,11 @@ Milestone 0 preserves:
 - Latest-generation-diff page schema version 1. Pages use one top-level pagination object and a
   nested `receipt_header`. `stored_receipt_hash` names the complete stored receipt. Opaque cursors
   may be restarted after a server or receipt change and are not durable public identifiers.
+- Manual render-plan schema version 1.
+- Graph view-plan schema version 1.
+- Projection-package schema version 1.
+- Projection-receipt schema version 1.
+- Independent projection-policy schema version 2. Effective policy version 1 remains frozen.
 
 Indexes, attestations, extraction caches, previews, and rendered artifacts are disposable. A schema
 change may rebuild them. Canonical project content and stored proposals may not be silently
@@ -161,12 +179,24 @@ The following guarantees remain stable:
 
 ## Rendering and visualization
 
-The `generic_html` renderer remains the supported version-1 manual projection. It retains confined
-paths, raw-HTML suppression, fixed template tokens, deterministic identities, atomic replacement,
-and side-effect-free status.
+The `generic_html` renderer remains the supported version-1 manual projection. Its public
+`GenericHtmlRenderer.prepare()` signature, renderer identity, frozen alpha bytes, confined paths,
+raw-HTML suppression, fixed template tokens, deterministic identities, atomic replacement, and
+side-effect-free status remain compatible. It now delegates through a versioned manual plan,
+immutable package, and independent renderer.
+
+The `portable_graph_html` renderer and `graph_render` descriptor table are additive. Manual and
+portable graph declarations, plans, policies, publication receipts, and status remain separate.
+The portable renderer does not replace the existing live viewer or `docforge_visualize`.
+
+Existing project descriptors may retain any positive `max_render_bytes` accepted by schema version
+1. A value above 20,000,000 bytes does not make the descriptor invalid, and a smaller actual
+artifact still renders. Actual detached worker transfer is a separate fixed 20,000,000-byte
+runtime boundary.
 
 The live graph viewer remains a read-only consumer of a generation-pinned validated index. It does
-not become project authority or MCP retrieval authority.
+not become project authority or MCP retrieval authority. Source reads use the pinned index
+generation instead of reopening mutable canonical files behind that generation.
 
 ## Task-context compatibility
 
@@ -199,8 +229,9 @@ The exact version-1 relation aliases are frozen by the MCP contract and reposito
 Changing an alias category requires a new planner version; it is not a silent implementation
 detail.
 
-`ManualRenderPlan`, `GraphViewPlan`, a portable graph renderer, and independently packaged
-renderers are later-milestone direction. Milestone 0 does not claim that those contracts exist.
+Milestone 3 adds `ManualRenderPlanV1`, `GraphViewPlanV1`, projection package and receipt version 1,
+and projection policy version 2. These are additive submodule and schema contracts. They do not
+change the legacy task-context, adapter, changeset, or effective-policy contracts described above.
 
 ## Safety boundary
 
@@ -221,6 +252,8 @@ Milestone 0 records rather than redesigns these areas:
   callers use targeted retrieval for that node.
 - One individually oversized changeset diff is transported as reconstructable canonical-JSON
   chunks. Cursors are corruption-detecting read tokens, not authenticated authorization tokens.
-- Manual planning is not separated from rendering.
-- There is no portable graph-planning or graph-rendering contract.
+- Production fragment validation is currently slower than forced-full rendering at the maintained
+  1,000-page fixture. Full rendering remains the equivalence and recovery oracle.
+- Remote render services, render farms, third-party renderer ecosystems, and a separate render MCP
+  remain deferred.
 - DocForge2 does not self-host its bootstrap documentation.
diff --git a/docs/CONTRACT.md b/docs/CONTRACT.md
index ec2b062..cfbaf30 100644
--- a/docs/CONTRACT.md
+++ b/docs/CONTRACT.md
@@ -24,6 +24,11 @@ commit when Git is available; it cannot change repository state.
 - Latest generation-diff page: `schemas/generation-diff-page.schema.json`, version 1.
 - Generated client configuration: `schemas/client-configuration.schema.json`, version 1.
 - Client doctor result: `schemas/doctor-result.schema.json`, version 1.
+- Manual render plan: `schemas/manual-render-plan.schema.json`, version 1.
+- Graph view plan: `schemas/graph-view-plan.schema.json`, version 1.
+- Projection package: `schemas/projection-package.schema.json`, version 1.
+- Projection receipt: `schemas/projection-receipt.schema.json`, version 1.
+- Independent projection policy: `schemas/projection-policy.schema.json`, version 2.
 - Index schema: version 3, disposable and reproducible.
 - Index attestation: schema version 1, disposable and reproducible.
 - Core, CLI, and MCP server: version 1.3.0.dev0.
@@ -82,7 +87,9 @@ the header alone. One top-level pagination object carries the only continuation
 Generated Codex, Claude, and OpenClaw fragments are machine-local projections. They are not
 canonical project content. Version 1 binds the selected project, exact isolated Python
 interpreter, canonical argument layout, effective policy, no-AST projection, render policy,
-timeouts, artifact bytes, and configuration hash.
+timeouts, artifact bytes, and configuration hash. Milestone 3 adds the version-2 projection policy,
+its hash, projection availability, and the exact descriptor hash to that attested configuration
+evidence. Omitted default selectors are recomposed against the bound descriptor.
 
 Preview is side-effect free. Explicit publication creates only one new private standalone
 fragment in an existing real directory. It never merges or replaces different content. Descriptor,
@@ -135,10 +142,12 @@ preview root, and one or more stable view IDs. Each view names a built-in render
 derived output file, title, and optional family filter. Paths are resolved under the project root
 and may not overlap canonical content, authority files, changesets, templates, or previews.
 
-The initial `generic_html` renderer uses pinned CommonMark parsing with raw HTML disabled. Templates
-are UTF-8 files with a fixed token vocabulary; they cannot name commands, modules, or executable
+The `generic_html` renderer uses pinned CommonMark parsing with raw HTML disabled. Templates are
+UTF-8 files with a fixed token vocabulary; they cannot name commands, modules, or executable
 renderers. Render identity covers the canonical source hash, optional changeset hash, selected node
-and edge identities, view configuration, template hash, renderer contract, and exact parser version.
+and edge identities, view configuration, template hash, renderer contract, and exact parser
+version. The frozen version-1 API and alpha bytes are preserved by a compatibility wrapper over the
+manual plan/package/renderer path.
 
 An explicit CLI render atomically replaces one declared derived output. MCP can render a validated
 changeset only to its isolated preview path. Normal status verifies bounded source, configuration,
@@ -146,6 +155,56 @@ template, output, renderer, and publication-receipt identities without reconstru
 Explicit deep status remains the side-effect-free full-render oracle. Input changes detected before
 atomic replacement fail without publishing a current receipt for stale output.
 
+## Independent projection boundary
+
+Manual and portable graph plans are separate version-1 contracts over one immutable validated
+generation. They use canonical JSON, deterministic ordering, fixed structural and serialized-size
+bounds, and content-derived identities. Plans contain selected graph facts and bounded content.
+They contain no live project object, database handle, absolute project or index path, arbitrary
+query, command, executable path, or caller-selected module.
+
+Projection packages bind one plan to inert assets, a closed built-in renderer identity, declared
+component versions, and an artifact inventory with a byte allowance. Receipts bind the exact
+package, plan, renderer, artifact hashes and sizes, diagnostics, timing, and detached peak memory.
+Manual and graph renderer modules accept only validated packages. They cannot select nodes, invent
+relationships, read project state, choose publication paths, or write canonical files.
+
+Detached execution uses one fixed private Python module, isolated mode, a trusted working
+directory, a sanitized environment, exactly one canonical newline-terminated JSON request and
+response, a closed renderer allowlist, a 30-second timeout, disk-spooled stdout, and bounded reads. The package
+contract is capped at 24,000,000 bytes and actual detached artifact transfer at 20,000,000 bytes.
+Project descriptors may retain a larger `max_render_bytes` compatibility allowance, but an actual
+detached transfer above the fixed worker boundary fails closed.
+
+Portable graph configuration is independent of manual render configuration. One view selects
+either an exact root or a bounded metadata-only lexical query plus closed filters and node, edge,
+depth, and work limits. Logic is excluded. The renderer emits a complete static Nodes, Flow, or Web
+artifact and uses JavaScript only as progressive enhancement.
+
+Portable publication commits a content-addressed artifact, renderer receipt, and one bounded
+generation/view manifest in that order. The manifest is the publication commit. Status reads only
+bounded manifest and receipt evidence and never plans or renders. Repair restores declared output
+only from validated content-addressed evidence. A failure after a replacement that cannot be
+proven rolled back returns explicit degraded committed evidence.
+
+Manual fragment records are disposable semantic cache entries. Their keys bind the renderer,
+component version, and complete page semantics. The detached renderer recomputes the expected page
+fragment before accepting cached bytes. Cold creation is compared with a full detached render
+before cache publication. Invalid, corrupt, forged, stale, individually oversized, or
+aggregate-oversized records fall back to the full oracle. The dedicated cache retains only current
+keys and is capped at 10,000 entries and 64,000,000 bytes.
+
+Projection policy version 2 composes manual `auto|explicit|disabled`, portable graph
+`explicit|disabled`, and live viewer `on-demand|disabled` independently. Active plan, render,
+application, onboarding, and viewer-start operations enforce the relevant policy before hidden
+work. Receipt-only status and explicit viewer stop remain available. Effective policy version 1
+and its legacy projections remain unchanged.
+
+The live viewer remains separate from portable graph publication. It consumes one
+generation-pinned validated index through the viewer manager. Source reads come from that pinned
+generation and do not reopen mutable canonical files behind an older snapshot. Neither live nor
+portable visualization is retrieval or canonical authority.
+
 Normal MCP access does not expose canonical application. An explicitly configured canonical
 applier registers one hash-bound application tool. No MCP mode exposes arbitrary renderer
 execution, arbitrary file writes, shell commands, Git mutation, build commands, deployment, or
diff --git a/docs/MILESTONE_3_BASELINE.md b/docs/MILESTONE_3_BASELINE.md
new file mode 100644
index 0000000..6c95702
--- /dev/null
+++ b/docs/MILESTONE_3_BASELINE.md
@@ -0,0 +1,90 @@
+# Milestone 3 baseline
+
+## Scope and method
+
+This baseline records the independent-projection behavior completed in Milestone 3. It was
+captured on 2026-07-29 from clean candidate commit
+`f5dccb5e1c312121f1af63780162f593d9363b98`.
+
+The maintained command was:
+
+```bash
+.venv/bin/python tools/milestone3_benchmark.py \
+  --mode full \
+  --nodes 1000 \
+  --samples 10 \
+  --output benchmarks/milestone3-2026-07-29.json
+```
+
+The synthetic generic fixture contains 1,000 manual pages, 1,000 portable-graph nodes, and 999
+edges. Durations use `time.perf_counter_ns()` and nearest-rank p95. In-process peak memory uses
+`tracemalloc`; detached worker peak memory comes from the worker receipt and `RUSAGE_SELF`.
+Every measured result is checked for deterministic semantic identity and bounded response size.
+
+Environment:
+
+- Linux 7.1.3-200.nobara.fc44.x86_64.
+- CPython 3.14.6.
+- x86_64.
+- Ten samples except the one-time production cold render and fragment-cache population.
+- In-process and detached-worker memory ceiling: 268,435,456 bytes.
+- Detached artifact-transfer ceiling: 20,000,000 bytes.
+
+The complete machine-readable result is
+[`benchmarks/milestone3-2026-07-29.json`](../benchmarks/milestone3-2026-07-29.json).
+
+## Results
+
+| Operation | Median | p95 | Limit | Maximum response |
+|---|---:|---:|---:|---:|
+| Manual full plan/package/render | 801.948 ms | 810.490 ms | 15,000 ms | 664 B |
+| Manual detached worker | 591.317 ms | 599.394 ms | 20,000 ms | 667 B |
+| Fragment-assisted equivalence | 638.999 ms | 666.119 ms | 15,000 ms | 664 B |
+| Production incremental cold | 2,827.152 ms | 2,827.152 ms | 20,000 ms | 668 B |
+| Production incremental warm | 2,143.388 ms | 2,206.540 ms | 20,000 ms | 669 B |
+| Production forced full | 944.135 ms | 978.870 ms | 20,000 ms | 668 B |
+| Fragment cache miss sweep | 106.203 ms | 106.888 ms | 5,000 ms | 145 B |
+| Fragment cache hit sweep | 498.911 ms | 514.324 ms | 5,000 ms | 145 B |
+| Portable graph full plan/package/render | 303.736 ms | 323.690 ms | 10,000 ms | 657 B |
+| Portable graph detached worker | 265.448 ms | 268.428 ms | 20,000 ms | 660 B |
+| Manual receipt-only status | 62.881 ms | 111.381 ms | 500 ms | 1,542 B |
+| Portable graph receipt-only status | 58.467 ms | 59.331 ms | 500 ms | 879 B |
+
+The largest traced in-process peak was 35,160,716 bytes. The direct manual worker track peaked at
+88,580,096 bytes and the portable graph worker at 89,583,616 bytes. The production manual paths,
+including cold, warm, forced-full, and mutation variants, peaked at 104,771,584 bytes. Every child
+peak was validated from its projection receipt against the 268,435,456-byte gate.
+
+The manual artifact was 583,149 bytes. The portable graph artifact was 718,383 bytes. The manual
+plan was 1,006,393 bytes and its ordinary package was 1,007,297 bytes. The graph plan was 398,158
+bytes and its package was 398,715 bytes.
+
+## Equivalence and no-work gates
+
+The benchmark proved exact output equivalence for:
+
+- Manual in-process and detached rendering.
+- Manual full and fragment-assisted rendering.
+- Production cold, warm, and forced-full rendering.
+- Production add, change, delete, and reorder variants.
+- Portable graph in-process and detached rendering.
+
+Manual and portable-graph status each performed zero project loads, source parses, adapter
+projection loads, adapter extraction, index checks, synchronization, index builds, render
+preparation, output construction, output hashing, and viewer-manager requests. Each status path
+performed only two cheap source-generation checks and verified committed receipt or manifest
+evidence.
+
+## Measured limits and future notes
+
+- Fragment reuse is a correctness, isolation, and recovery boundary in this milestone. At 1,000
+  pages, production warm fragment validation is slower than the forced-full path. Later
+  optimization must start from this measurement and preserve byte equivalence.
+- Full rendering remains the oracle and recovery path. Invalid, corrupt, oversized, stale, or
+  mismatched fragment records fall back without changing canonical facts.
+- The benchmark main-process `ru_maxrss` value was 102,692 KiB. It is cumulative across all
+  main-process operations and is recorded only as diagnostic context. Detached child peaks are
+  measured separately. Per-operation traced peaks and every detached receipt peak own the memory
+  gates.
+- The results do not justify a storage rewrite, render farm, remote renderer, or separate render
+  MCP.
diff --git a/docs/MILESTONE_3_CLOSEOUT.md b/docs/MILESTONE_3_CLOSEOUT.md
new file mode 100644
index 0000000..c46c193
--- /dev/null
+++ b/docs/MILESTONE_3_CLOSEOUT.md
@@ -0,0 +1,89 @@
+# Milestone 3 closeout
+
+## Outcome
+
+Milestone 3 is complete. Manual compilation, portable graph rendering, and the live viewer are
+separate generation-pinned consumers of the validated graph. They cannot become canonical or
+retrieval authority.
+
+Implemented contracts:
+
+- Version-1 `ManualRenderPlan`, `GraphViewPlan`, projection package, and projection receipt.
+- Strict canonical JSON identities and packaged Draft 2020-12 schemas.
+- Independent manual and portable-graph renderer import boundaries.
+- One isolated, fixed, one-request detached worker protocol with bounded request, response,
+  artifact, timeout, environment, and renderer inventory.
+- Content-addressed portable graph artifacts, renderer receipts, generation/view manifests,
+  receipt-only status, repair, and degraded committed-publication evidence.
+- Disposable semantic fragment records with bounded cache inventory, corruption recovery, and
+  full-render equivalence.
+- Version-2 independent projection policy while preserving version-1 effective-policy behavior.
+- Generation-pinned live source reads and a separate read-only viewer-manager lifecycle.
+- Automated axe-tag and keyboard gates for the manual, portable graph, and live viewer.
+- Repository-native contract, smoke, scale, response-size, memory, and equivalence gates.
+
+## Candidate evidence
+
+The frozen implementation candidate is
+`f5dccb5e1c312121f1af63780162f593d9363b98`.
+
+The complete repository gate passed:
+
+- Ruff formatting and lint.
+- HTML, rendered-manual HTML, portable-graph HTML, CSS, and JavaScript checks.
+- Pyright with zero diagnostics.
+- Warning-strict compilation and tests.
+- 281 tests and 272 subtests.
+- Three Playwright and axe accessibility flows. The alpha manual is checked with WCAG 2.0/2.1
+  A/AA axe tags; portable and live graph flows add WCAG 2.2 A/AA tags and keyboard interaction.
+- Lock and npm dependency-tree checks.
+- Wheel and source-distribution builds.
+- Milestone 0, 1, 2, and 3 smoke benchmarks.
+
+The maintained projection contract subset passed 142 tests and 236 subtests. A 10,000-node deep
+chain and one 10,000-node strongly connected component prove that manual cycle planning has no
+recursion-depth failure.
+
+An isolated wheel installation passed CLI and MCP startup, a real detached manual render, and the
+closed malformed-worker-request contract. Six Milestone 3 commits and the complete candidate tree
+passed Gitleaks 8.30.1 with no findings.
+
+Three independent adversarial review tracks covered manual isolation and fragment integrity,
+portable publication and policy binding, and worker/accessibility/benchmark gates. Reproduced
+project import, hostile environment, unbounded stdout, fragment forgery, cache growth, aggregate
+overflow, coordinated policy drift, render-limit compatibility, deep-graph, and module-startup
+defects were fixed and regression-tested before closeout.
+
+The clean ten-sample 1,000-node benchmark passed every threshold. Exact measurements, equivalence
+results, memory peaks, and response sizes are recorded in
+[`MILESTONE_3_BASELINE.md`](MILESTONE_3_BASELINE.md) and
+[`benchmarks/milestone3-2026-07-29.json`](../benchmarks/milestone3-2026-07-29.json).
+
+## Preserved boundaries
+
+- The `docforge` distribution, package, CLI, MCP executable, and existing tool names remain.
+- The frozen alpha manual remains exactly 2,043 bytes with its legacy output hash and render
+  identity.
+- Legacy one-method `load_projection()` adapters remain supported.
+- Effective policy version 1, no-AST behavior, and existing client bindings remain compatible.
+- Project descriptor schema version 1 and SQLite index schema version 3 remain unchanged.
+- Configured `max_render_bytes` values above the detached transfer ceiling still load; a small
+  actual artifact renders normally. Actual detached transfer remains capped at 20,000,000 bytes.
+- No storage replacement or self-hosting dependency was introduced.
+- WorldForge and ScrapeStation were not touched.
+- No production MCP integration was repointed.
+- The legacy Forgejo repository and `legacy` remote were not changed.
+- No tag, release, release announcement, or visibility change was created.
+
+## Known follow-up work
+
+Milestone 4 remains directional and is not active. Its adapter SDK and product-documentation work
+must not silently absorb these separate future ideas:
+
+- Optimize production fragment reuse only from measured profiles while preserving the forced-full
+  oracle.
+- Add authenticated cursors only if a stronger threat model requires them.
+- Verify Claude's native timeout representation.
+- Add versioned adapter-owned launcher metadata before generating custom-adapter configurations.
+- Keep remote render services, shared render farms, third-party renderers, storage replacement,
+  and self-hosting deferred until their own evidence justifies them.
diff --git a/docs/USER_MANUAL.md b/docs/USER_MANUAL.md
index a92ed1b..640f9ae 100644
--- a/docs/USER_MANUAL.md
+++ b/docs/USER_MANUAL.md
@@ -27,6 +27,10 @@ incremental methods while retaining the full loader as a fallback.
 - Opt-in incremental adapter extraction with reverse-dependency invalidation.
 - Lazy function-scoped logic projections that do not densify the primary graph.
 - Declared HTML render views. Arbitrary templates, render commands, and output paths are rejected.
+- Separate versioned manual and portable graph plans, immutable packages, detached built-in
+  renderers, and validated receipts.
+- Content-addressed portable Nodes/Flow/Web artifacts with receipt-only status and repair.
+- Independent manual, portable-graph, and live-viewer policy.
 - A loopback-only graph browser with Nodes, semantic Flow, and convergence Web views,
   relationship keys, source inspection, branch-aware node hiding, panel resizing, zooming, and
   managed idle shutdown.
@@ -47,7 +51,11 @@ Canonical files own facts:
 canonical Markdown/TOML or adapter sources
     ↓ validate
 disposable SQLite graph
-    ↓ query / visualize / compile context
+    ├── query / compile context
+    ├── ManualRenderPlanV1 → detached manual renderer → declared manual
+    ├── GraphViewPlanV1 → detached graph renderer → portable Nodes/Flow/Web artifact
+    └── pinned index → managed live Nodes/Flow/Web/Logic viewer
+    ↓
 people and agents
     ↓ propose
 isolated changeset + preview
@@ -77,8 +85,8 @@ source format.
 Clone and verify DocForge:
 
 ```bash
-git clone forgejo@repo.andraxion.net:administrator/DocForge.git /absolute/path/DocForge
-cd /absolute/path/DocForge
+git clone forgejo@repo.andraxion.net:administrator/DocForge2.git /absolute/path/DocForge2
+cd /absolute/path/DocForge2
 uv sync --group dev
 npm ci
 
@@ -159,6 +167,27 @@ output = "Docs/Rendered/Manual.html"
 title = "My Project Manual"
 families = ["architecture", "system", "operations", "roadmap"]
 
+[graph_render]
+output_root = ".docforge/portable-graph"
+
+[[graph_render.views]]
+id = "architecture"
+renderer = "portable_graph_html"
+output = "architecture.html"
+title = "Architecture"
+root = "architecture.overview"
+initial_mode = "web"
+depth = 3
+max_nodes = 250
+max_edges = 1000
+max_work = 100000
+families = ["architecture", "system"]
+relations = ["depends_on", "owns", "calls", "reads", "writes", "tested_by", "relates_to"]
+authorities = []
+statuses = ["current", "active", "verified"]
+tags = []
+include_logic = false
+
 [graph]
 allowed_relations = ["depends_on", "owns", "calls", "reads", "writes", "tested_by", "relates_to"]
 
@@ -492,13 +521,25 @@ builds or repairs the index.
 ```text
 render-status [VIEW_ID] [--deep]
 render VIEW_ID
+graph-plan VIEW_ID
+graph-render VIEW_ID
+graph-render-status [VIEW_ID]
 preview CHANGESET_ID VIEW_ID
 apply CHANGESET_ID --changeset-hash SHA256 --applier WRITER_ID
 ```
 
+`graph-plan` validates and returns one declared `GraphViewPlanV1` without publishing. A portable
+view must select exactly one stable `root` or metadata-only lexical `query`. It may use only Nodes,
+Flow, or Web as `initial_mode`; portable version 1 excludes function-scoped Logic.
+
+`graph-render` explicitly publishes the declared static artifact, content-addressed renderer
+evidence, and generation/view manifest. `graph-render-status` verifies only bounded committed
+evidence and never plans or renders. Portable publication is a local CLI action.
+
 The CLI apply command supports the generic adapter. It verifies that the configured writer owns the
 changeset, applies the exact reviewed hash, rebuilds the index, checks it, and regenerates every
-declared render. It does not commit or push the result.
+declared manual render when manual policy is `auto`. It does not publish portable graphs, commit,
+or push the result.
 
 ### Viewer commands
 
@@ -549,6 +590,20 @@ declares the named writer, and application requires the same writer/applier iden
 `--no-ast` to preserve the no-AST binding. Generic CLI generation refuses project-owned adapters
 because it cannot safely reconstruct their composition.
 
+Select projection behavior independently:
+
+```bash
+docforge configure codex \
+  --project /absolute/path/MyProject \
+  --manual-render-policy explicit \
+  --portable-graph-policy disabled \
+  --live-viewer-policy on-demand
+```
+
+The generated version-1 configuration result carries an additive version-2 `projection_policy`,
+its hash, projection availability, and the exact descriptor hash. Omitted default selectors are
+validated against that descriptor rather than trusted as self-reported output.
+
 Inspect one configured client binding:
 
 ```bash
@@ -569,6 +624,54 @@ starts MCP, executes the configured command, synchronizes, builds, renders, star
 writes configuration. Claude timeout representation and client filtering that cannot be proved
 locally remain explicit warnings.
 
+## Independent projection behavior
+
+Manual and portable graph renderers consume immutable, path-free packages. A package binds one
+generation-pinned plan, inert assets, fixed component versions, a built-in renderer identity, and
+an exact artifact inventory. The detached child cannot select nodes, open the project or index,
+choose a publication path, execute project code, or mutate canonical facts.
+
+Child startup is fixed to isolated Python, a private module entrypoint, a trusted working
+directory, and a sanitized environment. One request and response use canonical newline-terminated
+JSON. The request, response, receipt, execution time, and disk-spooled stdout are bounded. Actual
+artifact transfer is capped at 20,000,000 bytes even when the descriptor retains a larger
+`max_render_bytes` compatibility value.
+
+Manual fragment records are disposable semantic cache entries. On a cold miss, DocForge performs a
+trusted full detached render, extracts candidate page fragments, and compares fragment-assisted
+output byte-for-byte before publishing records. On a warm hit, the worker recomputes each expected
+page fragment before accepting cached bytes. Corrupt, forged, stale, individually oversized, or
+aggregate-oversized records fall back to the full oracle. Fragment reuse is currently a correctness
+and recovery boundary, not a promised speedup.
+
+Projection policy version 2 is:
+
+```text
+manual:         auto | explicit | disabled
+portable_graph: explicit | disabled
+live_viewer:    on-demand | disabled
+```
+
+For ordinary CLI commands, place the corresponding global flag before the subcommand:
+
+```bash
+docforge --project-root "$PROJECT" --manual-render-policy disabled render manual
+docforge --project-root "$PROJECT" --portable-graph-policy disabled graph-plan architecture
+docforge --project-root "$PROJECT" --live-viewer-policy disabled visualize
+```
+
+An active operation blocked by policy returns `projection_policy_forbids_operation` before hidden
+work. Manual and portable receipt-only status remain available. Viewer status and explicit stop
+remain available when viewer start is disabled.
+
+A non-disabled projection also requires its declared configuration or runtime. Manual `explicit`
+requires manual render configuration. Manual `auto` additionally requires canonical application in
+the current operation or server capability. Portable graph `explicit` requires portable graph
+render configuration, and live viewer `on-demand` requires its runtime. An unavailable selection
+returns `projection_policy_unavailable` before work begins. In particular, ordinary CLI `render`
+operations cannot select manual `auto`; use `explicit`, or let a configured canonical `apply`
+operation own automatic regeneration.
+
 ## MCP usage
 
 Run one MCP server per project with absolute paths:
@@ -586,7 +689,10 @@ Select the session's declared surface explicitly when useful:
 ```bash
 docforge-mcp \
   --project-root /absolute/path/MyProject \
-  --capability-mode read
+  --capability-mode read \
+  --manual-render-policy explicit \
+  --portable-graph-policy explicit \
+  --live-viewer-policy on-demand
 ```
 
 Supported modes are `read`, `proposal`, `application`, and `operator`. Existing startup defaults
@@ -613,8 +719,9 @@ identity, not a command. The changeset creator, configured writer, and canonical
 
 Call `docforge_bootstrap` first. Its version-1 `session_contract` contains the fixed binding,
 current graph generation, effective policy, actual capabilities, render policies, prohibitions,
-and a recommended first operation. Workflow guidance does not recommend registration or
-application when those startup capabilities are unavailable.
+and a recommended first operation. The result also carries the independently composed version-2
+`projection_policy` and hash. Workflow guidance does not recommend registration or application
+when those startup capabilities are unavailable.
 
 Example MCP client configuration:
 
@@ -653,11 +760,16 @@ Example MCP client configuration:
 - `docforge_get_task_context`
 - `docforge_validate_project`
 - `docforge_render_status`
+- `docforge_graph_plan`
+- `docforge_graph_render_status`
 - `docforge_visualize`
 - `docforge_visualization_status`
 - `docforge_stop_visualization`
 - `docforge_get_generation_diff`
 
+MCP graph plan and status are read-only. MCP does not expose portable graph publication; use the
+explicit local `graph-render` CLI command.
+
 ### Proposal tools
 
 - `docforge_create_changeset`
@@ -976,8 +1088,29 @@ docforge --project-root "$PROJECT" render-status
 docforge --project-root "$PROJECT" render VIEW_ID
 ```
 
-Successful canonical apply regenerates all declared views automatically. A manual canonical edit
-requires reindexing and rendering.
+Successful canonical apply regenerates declared manual views only when manual policy is `auto`.
+A manual canonical edit requires reindexing and explicit rendering. Portable graph publication
+always remains a separate explicit CLI action.
+
+Portable graph publication has separate status and policy:
+
+```bash
+docforge --project-root "$PROJECT" graph-render-status
+docforge --project-root "$PROJECT" graph-render architecture
+```
+
+### `projection_policy_forbids_operation`
+
+The process was deliberately started with the relevant manual, portable-graph, or live-viewer
+operation disabled. Restart with an allowed selector after confirming that the integration should
+receive that capability. Status and explicit stop operations remain available as described above.
+
+### `projection_policy_unavailable`
+
+The selected non-disabled projection has no matching project configuration or runtime. Add the
+declared manual or portable graph render configuration, or make the live viewer runtime available,
+before selecting that mode. Manual `auto` also requires an operation or MCP server with canonical
+application enabled. Use manual `explicit` for a standalone CLI render.
 
 ### Descriptor changed after startup
 
@@ -992,8 +1125,11 @@ Run the complete release gate from the DocForge repository:
 make gate
 ```
 
-Use `make benchmark` for the historical Milestone 0 baseline and `make benchmark-m1` for the
-counter-gated 1,000-node warm-operation benchmark.
+Use `make benchmark` for the historical Milestone 0 baseline, `make benchmark-m1` for the
+counter-gated warm-operation benchmark, `make benchmark-m2` for agent workflow gates, and
+`make benchmark-m3-full` for the ten-sample 1,000-node projection, worker, fragment, status,
+equivalence, response-size, and memory gates. `make accessibility` runs the generated manual,
+portable graph, and live viewer axe and keyboard flows.
 
 Project-specific vocabulary, extraction rules, and serialization belong in the project adapter.
 Generic core behavior must remain deterministic, project-bound, and recoverable.