ci: delivery metrics (lead time, red time, change failure, empty bee runs) defined in specs/ci/delivery_metrics.t27 #6600

Item H of epic #6595: delivery metrics in the DORA style, plus why bee runs end empty.

Measured read-only on 2026-10-05 over the window 2026-09-28T18:57Z .. 2026-10-05T18:57Z (gh, the ruleset t27-master-protection, check runs on 428 first-parent master commits, /queen/public-board, /queen/status, /queen/public-activity, and every queen-* branch on the remote). Headline numbers:

• 442 PRs merged (63.1 per day). Lead time from open to merge, median / p90: bees 6 min / 18.5 h (n=153), agents 26 min / 7.7 h (n=276), owner 374 h / 394 h (n=5), all 18 min / 15.6 h.
• Required checks (validate, check-linked-issue, parse-ratchet): 0 h red on master. Non-required: 9 checks red more than 48 h of the 168 h window (Scorecard analysis and spec-guards for the whole window, Corpus ratchet 101.5 h, fpga-conformance 99.6 h, duplicate-bodies 84.9 h).
• 0 reverts; 40 of 442 merged PRs (9.0%) were named by a fix PR merged within 24 h.
• 58 open PRs, 19 (32.8%) older than 7 days.
• 417 of 745 queen-* branches touched in the window (56.0%) end with no commit of the bee's and no diff.

Boundary

specs/ci/delivery_metrics.t27, its generated gen/c/ci/delivery_metrics.c, and one docs/now/ entry with the measured table. No tri command, no new script, nothing in BrowserOS.

User Scenarios & Testing

• A reader asks "how fast do we merge, and how often do we break master?" and finds the definitions in one spec and today's numbers in its tests.
• A future report or dashboard reads the definitions (lead time, red time, change failure, empty bee run) and alert thresholds from the spec instead of restating them.
• Negative controls: a value one step past each threshold alerts; a cancelled or skipped run does not turn a check red; a fix PR 24 h + 1 min after the merge, or one that only says Closes #M, is not a change failure.

Requirements

• Define lead time, author kind (bee / agent / owner / bot), red time per check and for the required union, change failure, and an empty bee run.
• Alert thresholds: lead p90 > 24 h; required red > 60 min per 7 days; any check red > 48 h per 7 days; change failure > 15%; open PRs older than 7 days > 25%; empty bee branches > 20%; fewer than 1 merge per day.
• Tests use today's measured numbers. Only t27 plus generated code.

Success Criteria

• t27c test-report specs/ci/delivery_metrics.t27 on the t27c lab: FAIL 0, and a mutated measured constant gives FAIL 1.
• validate, check-linked-issue, parse-ratchet, own-language and Corpus ratchet green on the PR.