Gate scoped memory audits with exact conflict and lifecycle regression fixtures

[Coding-Dev-Tools]

Adds eight synthetic memory-audit regression fixtures to the existing offline Spec Crawl gate. Each fixture runs in both input orders and compares exact conflict pairs, lifecycle actions, retirement targets and stored orphan connectivity. Compact node responses are required, empty datasets cannot claim success, and failed memory contracts fail the gate independently of document scores.

This PR follows #262. The original Oct 13 snapshot remains in Git history; its runtime tree was reconciled onto the reviewed #260–#262 stack, preserving scoped reads, temporal remediation, bounded/truncated reports, explicit clarification persistence and graph scheduling safeguards. The final feature diff carries the audit gate and its provenance.

Validation: 8/8 memory contracts in both input orders; five eval-gate regressions; 169 focused spec, API, polling, renderer, evidence, MCP and portable-skill tests; 13 documentation/distribution contract tests; Ruff. Immutable evidence and chart links bind the measured source; remote CI qualifies the current head.

[Coding-Dev-Tools]

Ten independent internal reviewers completed a read-only review of exact head d0f77d13fdb40599525480e470d4be607c489563. The agents covered full-file inventory, docs/evaluation, graph/UI, release/CI, MCP contracts, security/privacy, and data contracts. This is a comment review, not an approval; no tests were run.

Findings to reconcile before merge:

• P2 — Privacy default: clarification answers are saved as procedural memory by default (save_as_memory=true); without a session ID, the write resolves to workspace scope. Make persistence opt-in or require explicit confirmation.
• P2 — Incomplete audits appear complete: cluster analysis and procedural crawl cap records/text without consistently returning eligible totals or truncation status. Spec Studio also presents sampled results and empty states as full-workspace health.
• P2 — Scope and timeline regressions: repo-backed session crawls can omit repository scope; the resolver no longer enforces one requested session scope for both nodes; supersede closes the old fact at wall-clock time instead of the retained fact's valid_from.
• P2 — Graph regressions: showUnlinked=true can still hide degree-zero nodes when minDegree>0; a zero-size viewport can produce a bad fit, and worker updates can replace a user-owned camera fit.
• P2 — Spec crawl cost: range checks linearly rescan code ranges per line/token, allowing quadratic work on crafted input within the character cap.
• P2 — Evaluation metric: the four-case “precision” gate checks missing expected labels more than false positives; its name/threshold may overstate what it validates.
• P2/P3 — MCP schema mismatch: node_b is nullable/optional in endpoint schemas but required by service code; memory_ids and mtypes limits are enforced only in service code.
• One reviewer flagged a possible missing caller-visibility filter in the cluster analyzer; another reviewer did not reproduce a cross-user read from the default listing path. Validate that boundary before treating it as a confirmed leak.

Several findings match safeguards already present in the newer open PR #262. Keep this preservation snapshot in draft and reconcile those hunks before merge.