Scope and design

Closes #175. O2 of #173 / #101, milestone Daemon observability feature.

Implement the complete reviewed runtime instrumentation backend, not only the final retry correction:

• Fixed metric catalog, finite label domains, counters/gauges/histograms, 15-second sampling and bounded one-minute history; explicit units, windows, reset and unavailable/null semantics.
• Actual job admission, execution, retry, recovery, provider transport, embedding, confirmed cache/vector commit, writer contention, maintenance and Admin request/SSE observation hooks.
• Typed, opaque execution correlations and bounded retained execution-health evidence. Observation loss is separate from execution health and cannot fabricate success or silently claim complete coverage.
• Nonblocking bounded sinks outside owner locks/leases/commit turns; restart/replay/coalescing do not double-count actual work.
• Confirmed registration/source/identity-resolution transitions, origin/generation qualification, bounded primitive capture and discard on failed persistence.
• Preserve accepted core retries beyond the diagnostic ordinal catalog, while recording qualified count-only observation loss instead of a false ready state.

Whole change: 93 files, 12,057 additions / 191 deletions relative to main. Contracts and implementation evidence are in #175 comments and the updated component/test/cache-maintenance documentation.

UX and rollout

Backend cohort only. Public shared API/IPC/CLI/MCP contracts are explicitly sequenced in #176, embedded Admin UI in #177, integrated rollout in #178. This PR does not add an incomplete public screen or promise full #101 completion.

No external telemetry, arbitrary labels, credential-bearing data, raw provider bodies, repository-document byte storage, cache migration, installed-service changes or release.

Verification

Exact reviewed head: 4e313073c987d4940de9d4b3bbc7589744c4c38a; base main 4a7ae1817a5e0a6f48810a9121e1adb20042ae07.

Author: focused regression/race repeated three times; go test -p 1 ./...; relevant six-package race gate; vet, product-docs and diff whitespace checks passed.

Independent reviewer: full accumulated six-criterion audit CLEAR, no P0–P2; full Go gate and fresh race gates for observability, servicectl, repositorydocs, cache, service and adminhttp passed. Original independent retry-omission probe passes unchanged under race three times after the correction.

Exact-head CI and independent Luna verification: run 38103152507, attempt 1, push, .github/workflows/ci.yml; required job 114363081860 and all steps passed, including frontend semantic tests and artifact guard.

Prior failed observations remain recorded on #175: two earlier subprocess timing assertions exceeded their unchanged budgets, later isolated/full reruns passed (cause not established); reviewer long temporary-root socket fixtures passed unchanged from a shorter root. No timeout/assertion weakening. The independently reproduced retry diagnostic omission was fixed, not dismissed as flaky.

Risk and recovery

Broad hook coverage is the principal integration risk. Observation is bounded and diagnostic-only; failure/loss remains visible without changing core work authority. No cache-schema migration. Reverting this merged increment restores O1 while leaving O3/O4 rollout unavailable.

#188 is a separately reproduced obsolete-job core mutation after source rebind, not introduced or fixed by this observer-only O2 work. #178 must explicitly resolve its rollout disposition.

Merge requires canonical exact-head PR validation, no unresolved blocking discussion, and green required CI. Issue closure additionally waits for green CI on the resulting main commit.