frontier-auto: context-doctor gains an optional cache-boundary (prompt-caching) audit

[chrono-meta]

Summary

Autonomous weekly frontier-auto proposal (no human online). Adds an optional Step 3.5 — Cache-Boundary Audit (prompt caching) to plugins/fh-meta/skills/context-doctor/SKILL.md, with supporting detail in SKILL_detail.md §CacheBoundary.

Signal: 2026-07-27 comment on issue #102 ("Frontier Digest Log") — signal source, item 5 ("Context Engineering as the Dominant Production Discipline"), citing arXiv:2603.09619 and appscale.blog. The digest explicitly proposed: "Concrete improvement candidate: add a cache-boundary audit step to /context-doctor."

What it adds: context-doctor previously only diagnosed context-window footprint (Steps 1–5). Nothing addressed provider-side prompt-cache billing — a distinct cost lever. Step 3.5 adds:

• A checklist (system-prompt byte-stability, dynamic content placement) for diagnosing cache-hit degradation.
• The digest's reported decision thresholds (input:output ratio >10:1 favors cache/context engineering; >50:1 prefix caching dominates), explicitly labeled UNCALIBRATED per this repo's own Instrument-Calibration doctrine, since both cited sources returned HTTP 403 to this session's fetch tooling and were not independently re-verified.
• Wiring into the environment table, explicit/NL triggers, and Done When (so the step isn't orphaned).
• A cross-reference from Step 5 noting the compression-vs-cache-invalidation tension.

Predicted impact: recorded in tracks/_meta/edit_manifest.yaml (gitignored, local) — sessions with a cache-miss cost symptom on a metered backend now get a diagnosis path that didn't exist before; low-frequency trigger expected since it fires only on a specific cost symptom.

Verification (FH 4-axis gate)

• Axis 1 (regression_guard.sh --staged): PASS, 0 M-tier / 0 S-tier.
• Axis 2 (steel-quench-style adversarial review, dispatched to an isolated opus agent): Wave 1 found 1 S-grade + 3 A-grade issues on the first draft (bare/uncalibrated threshold, orphaned step, unsourced mechanism claim, body-bloat convention mismatch). All four fixed inline; Wave 3 re-attack by the same isolated agent converged PASS, zero surviving S/A.
• Axis 3 (phantom-quench-style claim check, inline): 5/5 numeric + citation claims traced via literal grep hit to the declared local source (the fetched issue comment). The external hop (arXiv/blog originals) is Unreachable (HTTP 403 in this sandbox) and is honestly disclosed as UNCALIBRATED in the doc itself, not silently treated as Phantom.
• Axis 4: predicted-impact entry recorded in tracks/_meta/edit_manifest.yaml.

Marker: tracks/_meta/.axes_23_passed_claude_frontier-auto-2026-07-27_2026-07-27.marker (gitignored, local — floor-status: at-floor, axis2-model: opus).

Test plan

• Human review of Step 3.5's placement/wording in plugins/fh-meta/skills/context-doctor/SKILL.md
• Confirm the UNCALIBRATED framing is acceptable (sources 403'd, not independently re-verified)
• Re-fetch arXiv:2603.09619 / appscale.blog from an environment with access, if/when convenient, to upgrade the figures from UNCALIBRATED to verified

This PR is opened as a draft per the frontier-auto routine's standing instruction: no auto-merge, human review required before landing.

---

Generated by Claude Code