release: v1.6 hardening — bounded imports, fleet hygiene, fail-able eval gates

.claude-plugin/skill-assets.sha256

@@ -2,5 +2,5 @@
 94bfa06317a8fe6a6a7e204bb70c5abdc9e4bbc34d79dd6f8447a30140bc8b85  .claude-plugin/plugin.json
 055655db84af07561d002f0c69744313d8413c39f3e873f941f0fa0b1e76dc66  skills/engraphis-memory/references/CONVENTIONS.md
 62019760766ff472a76a0f81437898f39e3c1fe2631732b7b7733e50c1ad837f  skills/engraphis-memory/references/SCOPING.md
-96c8e9b9cee1b3cb43c4bef9e48c57ed92af707f7f7a5d28d73b1ac247d2f0c6  skills/engraphis-memory/references/TOOLS.md
+1f62ba2b6abf3dab266d5b4d9c85f2d7fd7fe4ece4e85e2413cfc3fe460ef2c1  skills/engraphis-memory/references/TOOLS.md
 0f98098df695b9a00dc78402911124ebf09a4a058f6c8bec2c6234ec61fac13a  skills/engraphis-memory/SKILL.md

.github/dependabot.yml

@@ -26,3 +26,9 @@ updates:
     schedule:
       interval: "weekly"
     open-pull-requests-limit: 3
+    groups:
+      # codeql-action's init and analyze steps must always bump together —
+      # a mixed-version pair fails CodeQL's analyze post-action step.
+      codeql-action:
+        patterns:
+          - "github/codeql-action/*"

.github/workflows/ci.yml

@@ -37,7 +37,7 @@ jobs:
       - name: Unit tests (full suite — extras-gated tests included)
         run: |
           python -c "import fastapi, httpx, mcp, multipart, pydantic, uvicorn"
-          ENGRAPHIS_INDEX_ROOTS="${GITHUB_WORKSPACE}:${RUNNER_TEMP}" python -m pytest -o addopts="" tests/ -q -rs --basetemp="${RUNNER_TEMP}/engraphis-pytest"
+          ENGRAPHIS_INDEX_ROOTS="${GITHUB_WORKSPACE}:${RUNNER_TEMP}:$(python -c 'import tempfile; print(tempfile.gettempdir())')" python -m pytest -o addopts="" tests/ -q -rs --basetemp="${RUNNER_TEMP}/engraphis-pytest"
       - name: Retrieval eval gate
         run: python -m eval.harness --dataset eval/datasets/sample.jsonl --k 5
       - name: Retrieval eval gate — CodeMem (coding-agent wedge, incl. conflict resolution)
@@ -48,6 +48,10 @@ jobs:
         run: python -m eval.reinforcement
       - name: Adversarial memory prompt-boundary gate
         run: python -m eval.adversarial_memory_security
+      - name: Grounded-recall decision gate
+        run: python -m eval.grounded
+      - name: Code-agent arm gate
+        run: python -m eval.code_arm
 
   typecheck:
     name: core + backends typecheck (Python 3.11)
@@ -110,9 +114,11 @@ jobs:
           python -m pip install --upgrade pip
           pip install numpy "pytest<9"
       - name: Unit tests (extras-gated tests skip; the core must pass)
-        run: ENGRAPHIS_INDEX_ROOTS="${GITHUB_WORKSPACE}:${RUNNER_TEMP}" python -m pytest -o addopts="" tests/ -q -rs --basetemp="${RUNNER_TEMP}/engraphis-pytest"
+        run: ENGRAPHIS_INDEX_ROOTS="${GITHUB_WORKSPACE}:${RUNNER_TEMP}:$(python -c 'import tempfile; print(tempfile.gettempdir())')" python -m pytest -o addopts="" tests/ -q -rs --basetemp="${RUNNER_TEMP}/engraphis-pytest"
       - name: Retrieval eval gate
         run: python -m eval.harness --dataset eval/datasets/sample.jsonl --k 5
+      - name: Retrieval eval gate — CodeMem (coding-agent wedge, incl. conflict resolution)
+        run: python -m eval.harness --dataset eval/datasets/codemem.jsonl --k 5
       - name: Ablation
         run: python -m eval.ablation
       - name: Reinforcement state-transition gate
@@ -178,7 +184,7 @@ jobs:
           python -m pip install --upgrade pip
           pip install -e ".[test]" pytest-cov
       - name: Coverage run (all extras-gated tests, tracked modules)
-        run: ENGRAPHIS_INDEX_ROOTS="${GITHUB_WORKSPACE}:${RUNNER_TEMP}" python -m pytest -o addopts="" tests/ -q -rs --basetemp="${RUNNER_TEMP}/engraphis-pytest" --cov=engraphis --cov-report=term-missing --cov-fail-under=60
+        run: ENGRAPHIS_INDEX_ROOTS="${GITHUB_WORKSPACE}:${RUNNER_TEMP}:$(python -c 'import tempfile; print(tempfile.gettempdir())')" python -m pytest -o addopts="" tests/ -q -rs --basetemp="${RUNNER_TEMP}/engraphis-pytest" --cov=engraphis --cov-report=term-missing --cov-fail-under=60
 
   hygiene:
     name: repo hygiene gate (no stray DBs/logs)

.gitignore

@@ -114,3 +114,16 @@ cookies.txt
 
 # uv lockfile (generated tooling, not a project dependency)
 uv.lock
+
+# Schema-migration flock lives next to the DB (engraphis/config.py _migration_lock);
+# regenerable runtime state like *.db itself. Held live while the server runs.
+.*.migration.lock
+
+# One-off diagnostic/report dumps that keep landing at the repo root — regenerable
+# command output, never package content (same policy as /_*.mjs above).
+/404_paths.txt
+/disk_report.txt
+/large_files.txt
+/stats_pm2.txt
+/venv_status.txt
+/r3.txt

AGENTS.md

@@ -49,6 +49,8 @@ python -m eval.harness --dataset eval/datasets/codemem.jsonl --k 5  # coding/con
 python -m eval.ablation                                             # vector-only vs hybrid
 python -m eval.reinforcement                                        # bounded retention trajectory
 python -m eval.adversarial_memory_security                          # prompt/graph boundary
+python -m eval.grounded                                             # grounded-abstain decision gate
+python -m eval.code_arm                                             # coding-agent arm gate
 pyright                                                             # core + backends typecheck
 
 # ── External benchmarks (real numbers need torch + the dataset; see eval/external.py) ──
@@ -89,7 +91,9 @@ python -m scripts.migrate_to_v2 --old engraphis_v1.db --new engraphis_v2.db
 
 `requires-python >= 3.9` (ruff targets `py39`). CI tests the NumPy-only core on 3.9, the full
 offline stack on 3.10–3.14, and Pyright on 3.11; dedicated jobs also exercise encryption and built
-artifacts. `.github/workflows/ci.yml` is authoritative when the matrix changes.
+artifacts, and further jobs run the coverage gate (`--cov-fail-under=60`), repo hygiene, the Pi
+extension, browser accessibility, and the Docker smoke. `.github/workflows/ci.yml` is
+authoritative when the matrix changes.
 
 ---
 

BENCHMARKS.md

@@ -11,7 +11,7 @@ For the locked operator sequence for a public canonical run, see
 Every exact public aggregate retained below comes from the checked-in, public-safe
 [`offline-fixtures-v1.json`](docs/benchmark-evidence/offline-fixtures-v1.json) artifact. Its
 SHA-256 is
-`c3a74f1770ad3f868f55261ba11680e2dadca30167082ac2cb6669f9e3bdfad2`, also recorded in the
+`0f60b0868444f676fe14c5f94d7db2c475e22669930c4d760881d0842eaa6800`, also recorded in the
 adjacent `.sha256` file. The artifact contains no raw questions, answers, prompts, customer data,
 or per-record content fingerprints.
 

CHANGELOG.md

@@ -7,7 +7,6 @@ All notable changes to Engraphis are documented here. Format loosely follows
 
 ### Changed
 
-
 - Direct black-hole children now receive compact, deterministic orbital lanes near the black
   hole instead of inheriting the farthest authored radius. Each lane keeps phase and painted
   clearance, while community-child planets remain in their local moving frame; oversized Galaxy
@@ -97,9 +96,28 @@ All notable changes to Engraphis are documented here. Format loosely follows
 - Added `docs/GRAPH_PERFORMANCE.md` documenting the two graph presentation profiles,
   worker layout, progressive rendering, and the 20,000-node / 200,000-relation safety
   ceilings.
+- Source-import manifest paging now uses keyset (cursor) pagination instead of OFFSET,
+  so concurrent writes during a source re-import can no longer skip or duplicate rows
+  mid-scan (PR #154).
+- Local file/folder imports now accept up to 1,500 files per batch (was 500), with the total
+  batch ceiling scaled to 750 MB so the average per-file allowance is unchanged; document-wizard
+  scanner ceilings move in lockstep.
+- Folder imports report truncation explicitly: a folder with more matching files than the
+  ceiling now warns and returns `truncated`/`matched_total`/`unreadable` fields instead of
+  silently importing an alphabetically-first slice that looks complete.
 
 ### Fixed
 
+- Importing more than 1,000 files through the dashboard no longer fails with "Internal Server
+  Error": wizard upload routes parse multipart forms under the advertised 1,500-file ceiling
+  instead of Starlette's hidden 1,000-part parser default, oversized batches return a clear 413,
+  and large vault uploads no longer trip the dashboard's 8 MB default body limit.
+- One unreadable or pathological file (locked, deep-nested JSON, concurrent writer) now degrades
+  to a per-file error instead of rolling back the entire import batch with a 500.
+- Document/Obsidian import jobs whose worker died with the process are marked failed on the next
+  status poll (`worker_lease_expired`) instead of reporting `running` forever.
+- Cloud-placeholder files (OneDrive Files-On-Demand) on Windows are hydrated and imported rather
+  than rejected as non-regular files; symlinks and junctions remain blocked.
 - Galaxy layout now packs each complete solar-system envelope before orbital seeding and keeps
   those envelopes separated with rigid carrier translations during live motion. Compact server
   targets can no longer stack large systems near the black hole, while local planet positions,
@@ -131,15 +149,28 @@ All notable changes to Engraphis are documented here. Format loosely follows
   to register, instead of replaying the same broken asset response.
 - Existing Galaxy preferences migrate only the retired `48` orbital-separation default to `60`;
   deliberate custom values, including Gravity `0`, remain unchanged.
+- Source-import hardening lands via separate PR #154: deterministic missing-item detection
+  now guards an unknown baseline instead of reporting spurious misses, denial-guard
+  supersession binds digests computed from the parsed record rather than raw input,
+  import-job finalization is generation-guarded so a stale worker cannot finalize over a
+  newer attempt, and the finalized-state check completes in constant time.
 
 ### Security
 
+
 - HTTP error responses in `vault.py` and `service.py` no longer echo user-controlled paths back
   to the client, preventing filesystem structure leakage (SEC-001).
 - Graph visibility SQL helpers now use parameterized queries instead of `repr(float)` string
   interpolation, eliminating a fragile SQL construction pattern (SEC-002).
 - The `pypdf` dependency floor is raised to `>=6.15.0` to address PYSEC-2026-3655 and
   PYSEC-2026-3656 (arbitrary code execution via crafted PDF objects).
+
+### Removed
+
+- The Hermes memory-provider plugin integration (`integrations/hermes/`, its
+  `ENGRAPHIS_HERMES_*` environment surface, and its integration test) is withdrawn from
+  the repository ahead of the v1.6 tag. The provider remains available in the v1.5
+  release history for anyone who already copied it.
 ## [1.6] - 2026-08-15
 
 Minor release advancing the v2 engine through schema 16 with deterministic sync state, trusted