fix(skill): the report synthesis reads its baseline from the frontmatter, and three query scripts match their surface

[using-system]

Summary

• The report synthesis no longer names a baseline the run did not use. odd_report.py new writes a baseline: frontmatter field — the mission's named report (--baseline), else the recall's first match (same shared matcher as odd_recall.py, same-service-set rule), else none — and persist / show read it (or verifies on a replay). Older reports keep the prose fallback, which no longer takes a .md name from a line that opens with none / no previous.
• A re-measure's headline counts its baseline rulings, and the body contract names section 7's | Check | Before | After | Verdict | table.
• azure-monitor-traces.py exemplars ranks per operation and returns a p50 exemplar per operation (--slow N is now per operation, default 1).
• grafana-metrics.py labels retries once without the series an Adaptive Metrics rule aggregates, with a note.
• grafana-traces.py ops falls back to increase() of the span-metrics counter after a reset instead of withholding the calls.
• observe-run's zsh trap line gives one copy-pasteable sed -n 'A,Bp;C,Dp' form.

Choices amended against the issue text are recorded on the issue: #659 (comment)

Tests

• New tests in test_odd_report.py (baseline field from the recall / none / named / refused cases / same service set only, older-report prose guard, re-measure headline, Verdict column read by synthesis), rewritten exemplars test on re-captured masked fixtures, new Adaptive Metrics and reset-fallback tests for the grafana scripts.
• pytest tests/skills tests/hooks 1387 passed after the rebase on fix(agent): an environment hard stop persists nothing, and recall follows the status lineage #661; ruff 0.16.4 check/format clean; check_stack_reference.py green; apm install --target claude + apm audit green.
• show, synthesis and check over all 18 stored reports are byte-identical to main's; get-status --full identical.
• Live: Azure Monitor (exemplars p50/slow per operation) and Grafana Cloud (labels under Adaptive Metrics, ops after a real span-metrics reset) verified by the implementer and re-run by the reviewer.

Review

A separate reviewer sub-agent checked the whole branch under the bound-review rule:

• Round 1 found one blocking defect: a baseline the mission names was replaced by the recall's first line in the new field (a regression against main). Fixed with new --baseline (0 commands on the normal path, one contract line) and a test.
• Round 2: no finding. After the rebase on fix(agent): an environment hard stop persists nothing, and recall follows the status lineage #661 (the recall's equal-service-set rule carried into the shared matcher, fake-gcx env lists merged), round 3: no finding, ready to merge.

Harness measurement

test-plugin-harnessing, copilot / openai/gpt-5.6-luna / medium, /odd-observe drive mission on the local llms-benchmark stack, ABBA, both lab branches rebuilt right before the chain; every changed file in the deployed path.

Side (mean of 2)	Turns	Commands	Tokens in	Tokens out	Wall	Generation	Preflight	Observation
main	41.5	41.5	3.06 M	17.0 k	421 s	202 s	62 s	232 s
branch	31.5	39.5	2.30 M	15.4 k	320 s	139.5 s	32 s	158 s

Main's spread today over ten samples: turns 35–45, commands 29–54, tokens in 2.57–3.55 M, wall 310–431 s. No branch figure is above main's spread; several are below it (turns on both samples) — not claimed as a gain on two samples, this is a fix owing no degradation. 1 premium request per sample. Every sample persisted a report on the first try (branch: baseline: none and "no previous report" headline, consistent); no sample had a stored baseline, so the "vs baseline" path is covered by the unit tests only. An earlier chain was discarded (one main sample exited 1 at copilot shutdown, its replacement recalled a leftover report).

Closes #659

🤖 Generated with Claude Code