[EHotwagner]

A passive telemetry association kept at most 8 source cursors and 256 response identities and never retired either. On a host that writes about 150 Claude Code transcripts a day it deferred permanently after a few files. Measured capture-only on real 2.1.309 transcripts: two files filled the index in ten passes, and every later pass returned response-index-full-capture-deferred. This lands before the coherent113 candidate binds, so the fix ships in 0.113.0.

Change (SkillTelemetryAdapter.capturePassiveFromWithHook):

• Closing a source. A source closes when a pass reads its descriptor to the end on a line boundary, with no pass-budget, partial-tail or correction deferral, and the file has been unmodified for 15 minutes. PassiveScan gains SourceLength and ModifiedAtUnixSeconds. The cursor moves to a compact retired entry, and every response identity from that source instance leaves the index.
• Reselecting a closed source. A continuous append resumes from the retired entry. A rewritten file whose anchor no longer matches gets closed-source-discontinuity-skipped and is never re-read from zero.
• Active Claude transcripts keep only their newest 24 identities. The writer finishes all content blocks of one response before the next begins. Active Codex rollouts keep the original rule.
• Bound and horizon. At most 384 retired entries. Evicting the oldest sets a horizon one second after that file's modification time. Records from an unknown source at or before the horizon are suppressed (pre-horizon-record-suppressed) instead of being counted twice. Re-emitting an identity the receiver already holds would be refused as an identity conflict.
• Crash safety. Retirement rides the existing intent as absolute values (retirementAfter), so a crash at any cut applies it exactly once.
• Snapshot gains closedSources and retiredResponses.

Tests (WorkspaceTelemetryApplicationTests, 33/33; full FS.GG.Coord.Cli.Tests 754/754):

• 600 responses in one active session with no re-emission; the index holds 24.
• 20 files selected eight at a time: all close, resume reads only the appended responses, and a rewrite is skipped.
• Every crash cut during a closing pass: retirement is applied once and nothing is emitted twice.
• A three-day simulation: 451 files plus a main session growing each day, 1710 responses, each exactly once. State stays under 1 MiB, 384 retired entries, at most 256 identities. An evicted file reselected is suppressed, and a new file after the horizon is captured.

Also included:

• the 0.113.0 CLI release-note line
• roadmap paragraphs on retirement, re-read suppression and spool backpressure / drain cadence (utel-operational-completeness.md)
• the Atlas lesson github/passive-capture-finite-state-retirement (digest f5e33dfd…) in docs/atlas and both seeds, with the driver manifest regenerated

Review repair (after REQUEST CHANGES at 757a7c8b). Every re-read path of a retired response now suppresses instead of re-emitting, because the receiver would refuse a re-emitted identity as an identity conflict and quarantine the whole batch:

• source-alias-skipped: another active or retired entry already owns the same file (hard link, rename), meaning the same inode and the owner's anchored cursor still continues, so a reused inode is not taken as an alias. This also fixes a null-entry crash.
• continuity-replay-suppressed: an active file re-read after a continuity gap, at or before its high-water timestamp.
• retired-response-suppressed: an identity among the 4096 most recently retired (copies).
• closed-source-discontinuity-skipped: such an entry is now marked and not rescanned.

Quiescence takes the mtime after the read. The roadmap corrects three things: drain throughput is one envelope per passive chain per drain, settled receipts (not only applied ones) rotate, and the loss paths are listed. The Atlas record is replaced (now facb45d3…; earlier in-PR records never reached main). The release notes gain a rollback warning. After critic ACCEPT at 3b58ec3f: the alias check now requires anchored continuity, the replaced-inode test renames over the original to guarantee a new inode, the roadmap states the about-two-day copy window and the obligation to keep selecting active paths, and origin/main is merged. Tests: 34/34 targeted (a new alias/copy/rewrite regression fails on the previous adapter) and 755/755 for the full CLI suite.

Limits: the 24-identity window relies on observed Claude writer ordering, and a correction to an older response would surface as a receiver conflict. A single active Codex rollout with more than 256 responses still stalls until it closes. Installed adoption and Host ingestion remain separate.

Merge is held: the .github main merge hold (Wizard 0.22.0) is in effect, and the fdev sequence puts this before #4833.

architecture-map: unaffected

Refs #4640

🤖 Generated with Claude Code