[addisdev]

LOOP-PLAN Phase 1, all four items.

Summary

1. skipped is a job status. collector/src/skip.ts holds the one rule for how a job closes: ok: false on the final row is failed; a final row whose error begins skipped: is skipped; every device row skipped (none failed) under a plain ok: true host row is skipped; anything else is done. It is the same rule the v1 summary already used, so status and verdict cannot disagree.

• /results closes with it and writes the reason into last_error (for failures too, from the failing device when the host row has none).
• db.ts: a rebuild widens the status CHECK (copying whatever columns exist, by name) and, inside the same transaction, re-labels historical done jobs with a "skipped: result row. Runs once per database; result rows untouched; idempotent.
• depends_on: a waiter on a skipped job skips with dependency X skipped, cascading; a failure outranks a skip; enqueue behind an already-skipped job is skipped on arrival, and validate warns.
• Fan-out family status is skipped when no child ran (all skipped, or skipped + cancelled).
• /api/jobs and /api/v1/jobs status filters, schemas (gen:api regenerated), facets, /metrics, Overview skipped_24h, dashboard grey chip + filter option + "skipped 24h" stat, digest batch wait, and a jobs line in the morning digest. CLI/MCP verdicts are unchanged (they read the summary, which already said skipped).

2. schedule-skipping alert. One alert per enabled schedule whose last FLEET_ALERT_SKIP_STREAK (3) finished runs all skipped, e.g. nightly-tvloop-hardware has skipped 3 runs in a row: no Roku developer password on this host. Runs are found by job id (<schedule>-<YYYYMMDDHHMM>, fan-out children grouped) so pre-provenance jobs count; in-flight runs are ignored; resolves on the first non-skip.

3. Every failure says why. Diagnosis of the live case: job nightly-fleet-ui-smoke-202610150230's artifact 12de2ced… was uploaded and is in the store (in_store: true, 479-byte JUnit, served fine at /artifacts/<sha>). /api/artifacts/<sha> 404'd because no such GET route existed (only DELETE and /pin live there); the brain's catch-all answered "no such endpoint". Fix: GET /api/artifacts/:sha256 now 308s to /api/v1/artifacts/:sha256/content. The missing reason was the executor never reading the report: it contained Assertion is false: "Fleet Runner" is visible.

• junit.ts parses failing testcases; ui-test (Maestro) puts smoke: Assertion is false: … in error and test.failing_flow (added to result.schema.json). No report: Maestro's exit code + last stderr line, never the command line (it carries the test password).
• result-reason.ts, wired into fleet-client.postResult so every host workload gets it: a device row with ok:false and no error gets its test counts; a final host row with none gets the failing devices' reasons. The dispatch catch reports uncaught exec failures as exit code + last stderr line instead of Node's message.
• web-shots already says "N% of pixels differ" on its device rows; kept, and its final host row now inherits them.

4. Registry hygiene. (a) describeRoku and the fallback send platform: "roku"; os ~ 'roku' and tvloop (matches on the executor's Target.platform) unaffected. (b) A serial-named roku-* registration with descriptor.ip deletes roku-ip-<ip> (not when locked/claimed). (c) The lease sweep deletes rows whose descriptor declares kind or soc simulator and that have been offline > FLEET_SIMULATOR_TTL_DAYS (7); not real devices, not Android emulators, not the looser isSimulator() heuristic (it matches x86_64), not locked rows. POST /jobs/sweep returns forgotten.

Docs: concepts.md (new "Skipped is not done", chains), api.md, dashboard.md, deploy/alerts.md, deploy/index.md, protocol.md (platform list, skipped: convention, simulator TTL), lab.md, explore/findings.md (digest line), workloads/host.md (ui-test reasons), the lifecycle figure (re-rendered with npm run assets -- --only lifecycle), CHANGELOG [Unreleased].

Not verified

• Nothing ran against the live brain or real hardware; read-only GETs only, for the diagnosis. The migration/backfill is tested on a hand-built 0.8-shaped database, not a copy of the brain's.
• The Maestro "no report" path is unit-tested through execFailureReason, not with a real failing Maestro.
• The 0.7.2 brain will keep showing skips as done until Phase 2 upgrades it.
• Decision taken without asking: the morning digest still sends only when there are findings (the jobs line rides along); Android emulators are not forgotten. Easy to change if you want otherwise.

Test evidence (on ultra)

collector CI=true npm test:

ALL PASS

(1215 ok lines; new: skip: ×10, chain: skip ×5, fail: ×1, artifact: ×3, roku: registry ×3, sweep: ×1, skipped: migration ×11, stale skip: ×9, junit: ×9, exec: ×5, reasons: ×7, hygiene: ×10, Roku platform ×3, digest jobs ×3)

fleet CI=true npm test (with runner-machine deps installed so nothing skips):

ℹ tests 80
ℹ pass 80
ℹ fail 0
ℹ skipped 0

mkdocs build --strict: exit 0.

🤖 Generated with Claude Code

[addisdev]

Self-review: skipped status + idempotent relabel (verified on a copy of the live brain's database: 28 historical skips relabelled, one mixed job correctly left done), stale-skip alert, failure reasons from Maestro JUnit, Roku platform and registry hygiene. Merge conflicts with #92/#93 resolved (one duplicate test import fixed). All checks green. Merging (LOOP-PLAN Phase 1).