docs: complete frozen fork capability accounting for #100

[Skywind5487]

Summary

• Completes Complete #96 semantic capability accounting from frozen fork history #100 semantic capability accounting from the frozen feat: add reproducible fork archaeology inventory (#96) #99 inventory.
• Keeps the existing 145 change records + 26 merge events explicitly accounted (171/171).
• Adds an adversarial hunk/intent-level behavioral coverage audit so a commit can map to multiple semantic intents.
• Splits memory trim policy from memory trim diagnostics.
• Promotes three additional hidden behavior boundaries: compression no-op session boundary, SQLite write-contention backoff, and session-search context hydration/lock scope.
• Keeps provisional inventory buckets as provenance sources rather than capability boundaries; no archaeology framework changes and no dev rewrite/rebase.

Evidence

• Frozen fork: 35c8564c9c0af3d75bcbdf1d793e7207e5528f06
• Frozen upstream: 460d345642ee3d143a3e461abe39fd42b86a7e54
• Merge base: 91937a6dc3ffbbe2f3be91a500f0ecf962c4cf53
• Archaeology inventory: discovery PASS and explicit commit/merge accounting PASS (171/171).
• Behavioral audit: 145 unique change records, 26 merge events, 149 hunk/intent rows, 141 behavior-bearing rows, 5 promoted separate-intent rows, 0 uncovered behavior rows.
• Manifest: 26 reviewed groups: 23 provenance-backed groups plus 3 explicit behavioral overlays.

False-negative corrections

• 04f1af72 / a9d2b9af: threshold_mb and independent gc_cooldown_seconds are runtime policy, not only telemetry.
• 176646d2: compressor-reported noop skips session rotation/rewrite; kept separate from lifecycle telemetry.
• cc2531fb: SQLite retry changes to exponential 20ms-base / 2s-cap backoff; kept separate from incidental diagnostics.
• b879cbd3: bounded batch context hydration and lock scope are separately accounted from benchmark evidence.
• 7b66bbf8: benchmark cold-open/fallback checks are explicitly retained as validation evidence, not promoted as runtime capability.

Durable artifacts

• artifacts/fork-archaeology-issue-100/capability-manifest.md
• artifacts/fork-archaeology-issue-100/behavioral-coverage-audit.md
• artifacts/fork-archaeology-issue-100/behavioral-coverage-audit.json
• artifacts/fork-archaeology-issue-100/inventory.json and inventory.md

Verification

• scripts/fork_archaeology.py regenerated inventory artifacts successfully.
• 17 passed for tests/scripts/test_fork_archaeology.py and tests/scripts/test_capability_manifest.py via the repository pytest fallback.
• The canonical scripts/run_tests.sh could not start in this Windows environment because its Bash service returned E_ACCESSDENIED.

Review boundaries

This PR remains draft and does not merge, create fork/<feature> branches, modify/rewrite/rebase dev, or expand PR #99 tooling. Phase-2 port boundaries and evidence gaps remain recorded in docs/research/fork-capability-accounting-issue-100.md.

[Skywind5487]

Current-upstream ownership correction discovered from Phase-2 ticket #113

Phase-2's mandatory fresh SQLite prior-art pass found that the historical capability promoted here from cc2531fbc6df... is no longer a fork residual on the reconstruction substrate.

The frozen Phase-1 provenance remains valid: cc2531fb did introduce the fork's exponential 20ms-base / 2s-cap contention policy relative to the frozen comparison. The correction is to current ownership, not to the historical behavioral accounting.

Primary evidence:

• fix(state): time-based write-lock patience so busy sibling processes can't destroy turns NousResearch/hermes-agent#74480 merged as 8da8a7887d06373e169af6e431ec52ebb439ec7a, replacing attempt-counted retry with accepted time-based write-lock patience (20s routine / 60s transcript, fast 20–150ms jitter then 250ms–1s slow jitter, explicit exhaustion, non-lock propagation).
• That commit is already an ancestor of Phase-1.5 PR docs: compose Phase-1 capabilities into feature lines #108's pinned upstream 56526bc0d36522ab7a87ee0056f70e3847d2f0e6, so Phase 1.5 should have recorded this as PHASE1_DRIFT / current UPSTREAM_OWNED rather than projecting a PORT feature line.
• Upstream fix(state): four session-state fixes — safe close tracking, flush-cursor class fix, row-retry, usage-PK healer NousResearch/hermes-agent#75883 later merged the separate no more rows available retry class while explicitly preserving/superseding older retry constants with the 8da8a788 patience model.
• Current fork dev and current upstream main already use the accepted time-based writer policy.

Accordingly, keep this PR's historical capability/provenance record, but downstream/current-upstream composition must treat capability:sqlite-write-contention-policy as externally satisfied. See the blocking correction note on PR #108 and resolution of #113.