docs(skill): document oc_journal_compact cadence and pin round-trip stability

[shaun0927]

Summary

Stacked on #1442 (Part 1).

• docs/skills/trajectory-compaction.md: cadence table (20–50 step), strategy selection, composition with the LLM-free fast path (docs(skill): document LLM-free recall→precheck→replay fast path #1443).
• tests/tools/oc-journal-compact-roundtrip.test.ts: pins the lossy-but-stable contract — two consecutive compactions agree on open_assertions, last_checkpoint, and (deterministic strategies) the summary text; plus a sampling-unsupported case pinning the SSOT: OpenChrome as a host-neutral MCP browser harness #1359 "never server-side" guarantee.

Test plan

• Round-trip test 3/3 pass.

Review-pass changes

• Added the sampling-unsupported test (asserts status:'unsupported_by_host' with no fabricated summary).
• Softened the cadence table so the 20/50-step figures stay attributed to Webwright rather than implying an OpenChrome-measured benchmark (SSOT: OpenChrome as a host-neutral MCP browser harness #1359 P5).

Merge note: base branch is feat/1434-journal-compact (#1442). Landing this content on develop is gated on #1442, which is currently conflicting with develop.

Part 2 of #1434. Depends on #1442.

Co-Authored-By: Claude Opus 4.7 (1M context) noreply@anthropic.com

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.

[shaun0927]

Goal-review verdict — #1444

Intent vs SSOT (#1359)

• Pillar A (MCP portability) + Pillar D (portable verified memory) — docs-only addition that pins the cadence guidance and locks the lossy-but-stable contract via a round-trip test. No tool surface change, no new host behavior; the doc explicitly states OpenChrome never falls back to a server-side LLM for the sampling strategy. Aligned with P2 (harness, not agent) and P4 (facts before decisions).

Code-level review

• docs/skills/trajectory-compaction.md: cadence table is consistent with the empirical anchors in benchmark: episode budget shaper and early-stop using Webwright's marginal-utility curve #1428 (Webwright's "next 50 steps deliver only 3–4 additional accuracy points") and cross-links to docs/skills/llm-free-fast-path.md for composition.
• tests/tools/oc-journal-compact-roundtrip.test.ts: 2/2 pass locally — open_assertions, last_checkpoint, fact count, and (deterministic) summary text are byte-identical across two passes. This is the right invariant for "lossy but stable".

CI fixes applied (rebased on the updated parent)

• Rebased onto feat/1434-journal-compact head e756c47a (parent feat(tools): add oc_journal_compact deterministic core #1442) which carries the previously-missing TOOL_CAPABILITY_MAP entry and v1.11 snapshot row for oc_journal_compact. With those in place the inherited tests/capability-filter.test.ts failures clear.

Status

Merge-ready pending parent #1442 landing. No further changes to this PR's own diff.

[shaun0927]

Review analysis — merge-readiness

Intent & direction. Part 2 of #1434: documents the oc_journal_compact cadence/strategy and pins a round-trip stability contract (compaction is lossy-but-stable). Sound and matches the issue.

#1359 alignment — PASS. The docs correctly describe the sampling strategy as returning unsupported_by_host with no server-side fallback, and the implementation matches. The cadence figures are Webwright's, not OpenChrome benchmarks.

Findings & fixes applied this pass:

• (was P1) Untested SSOT: OpenChrome as a host-neutral MCP browser harness #1359 guarantee. The "never server-side" claim was prose-only. Added a third round-trip test asserting strategy:'sampling' → status:'unsupported_by_host' with no fabricated summary when the host lacks the capability (commit a8e838df).
• (was P2) Doc over-claim. Softened the cadence table so the 20/50-step figures stay attributed to Webwright (SSOT: OpenChrome as a host-neutral MCP browser harness #1359 P5).
• (P2) server as any test cast to reach the private tools map is acknowledged (eslint-disabled, author-aware); left as-is — low risk for an internal test.

Tests: oc-journal-compact-roundtrip.test.ts 3/3 pass locally; typecheck clean.

Merge logistics (the one real blocker for develop): base is feat/1434-journal-compact (#1442, Part 1), which is currently conflicting with develop. This PR is content-ready and is being merged into its base branch (the stacked-PR path); its content reaches develop only when #1442's conflict is resolved and #1442 merges.

Verdict: content MERGE-READY; landing on develop is gated on #1442.