Gate scoped memory audits with exact conflict and lifecycle regression fixtures

BENCHMARKS.md

@@ -94,14 +94,14 @@ interpretation and do not count as additional benchmark-quality gains.
 ### Public numeric evidence registry
 
 Every exact public aggregate retained below comes from the checked-in, public-safe
-[`offline-fixtures-pr262-git-bytes-20261014.json`](docs/benchmark-evidence/offline-fixtures-pr262-git-bytes-20261014.json) artifact. Its
+[`offline-fixtures-pr263-review-boundaries-20261016.json`](docs/benchmark-evidence/offline-fixtures-pr263-review-boundaries-20261016.json) artifact. Its
 SHA-256 is
-`1b4ba1d95ae003bab4c5d15a7d619fc7a7b813d66387bc9530a74babad5221eb`, also recorded in the
+`87fed662d344f5cce1640452f21058b5c5eb1e76a6e6df397378a0731182f18e`, also recorded in the
 adjacent `.sha256` file. The artifact contains no raw questions, answers, prompts, customer data,
 or per-record content fingerprints.
 
 The fixture-suite digest is
-`e5f78cc61a584b64051cc8650c8e2f68cff9a3ccb096df57e8a621d376f829b8`. The artifact defines
+`d10e174ef0768ffe70dd2d004cecd4d4a2eb6056c4f3a50bcff05b15b0b826ae`. The artifact defines
 the digest algorithm and records the SHA-256 of every suite and dataset file. Each evidence ID
 also binds its exact command through `sha256(UTF-8 exact command)`:
 
@@ -126,13 +126,13 @@ grounded checks in separate panels; provider billing and MCP transport are not m
 the SVG and matching PNG with:
 
 ```bash
-python scripts/render_benchmark_report.py --report docs/benchmark-evidence/offline-fixtures-pr262-git-bytes-20261014.json --output docs/images/context-efficiency.svg --png-output docs/images/context-efficiency.png
+python scripts/render_benchmark_report.py --report docs/benchmark-evidence/offline-fixtures-pr263-review-boundaries-20261016.json --output docs/images/context-efficiency.svg --png-output docs/images/context-efficiency.png
 ```
 
 The companion examples are also generated from that artifact with:
 
 ```bash
-python -m scripts.render_benchmark_examples --report docs/benchmark-evidence/offline-fixtures-pr262-git-bytes-20261014.json --output docs/images/evidence-backed-agent-examples.svg --png-output docs/images/evidence-backed-agent-examples.png
+python -m scripts.render_benchmark_examples --report docs/benchmark-evidence/offline-fixtures-pr263-review-boundaries-20261016.json --output docs/images/evidence-backed-agent-examples.svg --png-output docs/images/evidence-backed-agent-examples.png
 ```
 The historical-to-executable mapping is in
 [`docs/BENCHMARK_CHANGE_COVERAGE.md`](docs/BENCHMARK_CHANGE_COVERAGE.md).

README.md

@@ -50,7 +50,7 @@ Connect a coding agent over MCP with the [agent setup guide](https://github.com/
 The current registered artifact contains three deterministic offline fixture runs. It separates context size, retrieval quality, and grounded decision checks; these small fixtures do not establish general task performance.
 
 <p align="center">
-  <img src="https://raw.githubusercontent.com/Coding-Dev-Tools/engraphis/dfa3151d06ca3c226b50bc2e43305f4b23d42903/docs/images/context-efficiency.svg" alt="Three registered offline fixtures: structure-aware chunking reduced mean retrieved context from 740.3 to 214.3 tokens per question (71.1%, 18 questions); the serialized JSON-shape proxy fell from 24,590 to 11,138 tokens across 26 payload samples and 260 recalls. Candidate and packed retrieval quality (Recall@5, Hit@5, and answer-token recall) are shown separately (each 1.000). Grounded checks show 5/5 answerable queries grounded and 6/6 abstention queries rejected, including a 1/1 quarantined-evidence probe; 11/11 decisions were correct. MCP transport and provider billing were not measured. Artifact SHA-256 prefix 1b4ba1d95ae0; see the benchmark guide for the full checksum." width="100%">
+  <img src="https://raw.githubusercontent.com/Coding-Dev-Tools/engraphis/e574baa806eef561b24ac1318fb3abebb4432956/docs/images/context-efficiency.svg" alt="Three registered offline fixtures: structure-aware chunking reduced mean retrieved context from 740.3 to 214.3 tokens per question (71.1%, 18 questions); the serialized JSON-shape proxy fell from 24,590 to 11,138 tokens across 26 payload samples and 260 recalls. Candidate and packed retrieval quality (Recall@5, Hit@5, and answer-token recall) are shown separately (each 1.000). Grounded checks show 5/5 answerable queries grounded and 6/6 abstention queries rejected, including a 1/1 quarantined-evidence probe; 11/11 decisions were correct. MCP transport and provider billing were not measured. Artifact SHA-256 prefix 87fed662d344; see the benchmark guide for the full checksum." width="100%">
   <br>
   <sup>Three offline fixtures separate context reduction, candidate and packed retrieval quality, and grounded behavior. The chart shows the artifact checksum prefix; see the benchmark guide for the full checksum and reproduction steps.</sup>
 </p>
@@ -61,7 +61,7 @@ The current registered artifact contains three deterministic offline fixture run
 | Recall payload proxy | JSON-shape proxy: 24,590 → 11,138 tokens (54.71% lower, 26 samples; 260 timed recalls) | Candidate and packed Recall@5, Hit@5, and answer-token recall are each 1.000 |
 | Grounded decisions | 5/5 answerable queries grounded; 6/6 abstention queries rejected, including 1/1 quarantined-evidence check | 11/11 decisions correct |
 
-The payload figure is a serialized JSON-shape estimate, not an MCP transport measurement or provider billing total. See the [Benchmark methodology](https://github.com/Coding-Dev-Tools/engraphis/blob/dfa3151d06ca3c226b50bc2e43305f4b23d42903/BENCHMARKS.md) for artifact identity, counting methods, reproduction commands, external-evaluation boundaries, and limitations.
+The payload figure is a serialized JSON-shape estimate, not an MCP transport measurement or provider billing total. See the [Benchmark methodology](https://github.com/Coding-Dev-Tools/engraphis/blob/5f94cd2a1f01c5627cb62104015d184ca9c30f01/BENCHMARKS.md) for artifact identity, counting methods, reproduction commands, external-evaluation boundaries, and limitations.
 
 ## Optional Jev assistance
 

docs/DOCUMENT_IMPORT.md

@@ -107,6 +107,10 @@ rather than silently split. Containers are additionally capped at 2,000 members
 declared decompressed content. Invalid UTF-8/UTF-16 in permitted text is replaced explicitly and
 reported as a parsing warning.
 
+Library Quick import applies the same credential filename exclusions to uploaded files and
+folder paths before extraction or optional fact derivation. Hidden and dependency directories
+are skipped as well, and a symlink's target cannot bypass the filename exclusions.
+
 Markdown uses the Obsidian adapter's stricter 2 MB raw-note limit before decoding.
 
 Review a preview before importing sensitive material. The local importer makes no network request

docs/SPEC_CRAWL.md

@@ -96,6 +96,7 @@ The unified dashboard at `http://127.0.0.1:8700` includes a dedicated **Spec Stu
 Beyond inspecting isolated single-text prompt specifications, Spec Crawl provides cluster-level coherence auditing across multiple active memory records (`engraphis/core/spec_crawl.py::analyze_memory_nodes`):
 1. **Contradiction Detection**:
    - Parameter collisions for positively matching, nonempty subject keys (e.g. `port: 5432` vs `port: 5433` for the same database).
+   - Sentence-final periods on numeric port, timeout, and retry values do not create a parameter conflict; decimal values remain intact.
    - Opposing modal directives (e.g. `must enforce <target>` vs `never allow / forbid <target>`).
    - Divergences for the same `(subject_key, claim_kind)` are review suggestions; different keyed subjects do not establish parameter conflicts.
    - A keeper suggestion requires distinct, finite `valid_from` dates. Ingestion order never determines which fact is current. Missing, invalid, or equal effective dates require explicit clarification without a retirement target.
@@ -111,6 +112,10 @@ Beyond inspecting isolated single-text prompt specifications, Spec Crawl provide
 
 Memory bodies are used during analysis but are not returned in node summaries. The workspace audit analyzes at most 150 eligible memories per request and reports its eligible total and truncation state; Spec Studio warns when its health score and findings describe only a partial sample.
 
+Title aliases and body signals match whole words, so titles such as “Preview configuration” do not count as review policies. Similarity comparisons use at most 512 distinct tokens per memory, selected deterministically by token hash, and reuse each pair's result across link, orphan, divergence, and redundancy analysis. Reports expose `similarity_token_limit` and `similarity_sampled_node_count`; each node includes `similarity_token_count` and `similarity_sampled`. Sampled similarity is approximate and produces suggestions for review. Parameter and directive conflict checks inspect the complete memory prose outside fenced and inline code examples. Code examples are also excluded from prompt injection scoring.
+
+Word classification reads at most `word_limit` (5,000) tokens per memory. Each node reports `word_analysis_truncated` and `word_count_is_lower_bound`; a truncated `word_count` is the number analyzed, not the full memory length. Reports include `word_analysis_truncated_node_count`, and partial nodes receive no clean-node quality credit. HTTP and MCP audits set `truncated` and `analysis_truncated` for either omitted memories or incomplete word analysis; `node_selection_truncated` identifies an omitted-memory sample separately.
+
 ---
 
 ## 7. Command Line Interface (CLI)
@@ -170,12 +175,12 @@ Executes an explicit remediation on two live, approved memories after checking s
 
 ## 9. Smart MCP Tools
 
-The two read-only crawls are available to hosted/team viewers; clarification and resolution require member access.
+In hosted/team Classic MCP, the two read-only crawls allow viewer access; clarification and resolution require member access. In remote Smart MCP, crawls use the viewer-accessible `engraphis_execute_read`, while clarification and resolution use `engraphis_execute_action` and require admin access. Local stdio retains the operator's full capability.
 Omit `workspace` when passing a `session_id` to inherit its workspace and repository; explicitly mismatched scope selections are rejected. Default cluster audits return at most 150 eligible nodes and report `truncated`, `node_limit`, and `count_semantics`; the dashboard labels truncated health scores as a sample. Inferred similarity links are suggestions and do not count as stored connectivity or remove orphan remedies.
 
 These four Classic MCP tools are also discoverable and callable through Smart MCP actions:
 - **`engraphis_spec_crawl`**: Read-only prompt spec crawler returning scores, radar coverage, classified tokens, connection tentacles, and flags.
-- **`engraphis_spec_crawl_answer`**: Clarification answer submission updating specifications and writing pending procedural memories.
+- **`engraphis_spec_crawl_answer`**: Clarification answer submission updating specifications and optionally saving pending procedural memories.
 - **`engraphis_spec_crawl_memories`**: Audits active workspace memory nodes for contradictions, orphans, redundancies, and cluster health.
 - **`engraphis_spec_crawl_resolve`**: Applies an explicitly requested link or confirmed supersession to selected memory nodes.
 
@@ -185,6 +190,8 @@ These four Classic MCP tools are also discoverable and callable through Smart MC
 
 CI verifies determinism, axis classification, flag recall, qualified-token avoidance, section counts, and score boundaries in both the full stack and NumPy-only jobs. Qualified-token avoidance measures the explicitly labelled negative examples; it is not a general false-positive precision estimate:
 
+The same gate checks eight synthetic memory-audit contracts in both input orders: unrelated-subject noise, agreeing and opposing nested directives, subject and claim-kind separation, stored orphan connectivity, unknown effective dates, and effective-time ordering. It compares exact conflict pairs and retirement targets, requires compact node responses, and reports the fixture digest. These bounded regression fixtures do not establish general classifier quality.
+
 ```bash
 python -m eval.spec_crawl
 ```