[Juliusolsson05]

Fixes #678

What was wrong

entryClaims matched a queued task-notification on its task id alone. The id is present on every recorded notification, but it is not unique: upstream reuses it when reconciling a background shell from a previous session. In the recording, the pairs at events 111/112 and 359/360 are "…was stopped" followed by "No completion record was found…", each under one id. find returned the first twin, so it was retired as delivered-observed with the second twin's uuid as evidence, and the item that actually departed stayed pending.

Fix

resolveEntryTarget (in reconcile.ts) mirrors resolveRemoveCarrierTarget:

1. Prefer the claimant whose whole normalized body is contained in the committed entry. Containment, not equality, because entries wrap notifications.
2. Otherwise fall back to the id, but only when a single pending item carries it. Twins with no body match mean no delivery can be proven, so none is claimed.

Batched deliveries, where one entry carries several different ids, are unaffected.

Corrected expectations (evidence, not convenience)

Two corpus tests had blessed the misattribution. #678 predicted this.

• "uses the recorded remove content in a mixed queue": it retained event 112, the twin that departed at 114. It now retains 111.
• "does not strand background commands": it passed only because the first twins were retired with their siblings' evidence. In the recording, each pair has one dequeue and one committed entry, and that entry carries the second body; nothing shows the first leaving. They now stay pending and become unconfirmed on the idle sweep (the strip's existing honest state), and the test asserts exactly that.

Tests

• New, from recorded events 111–114: the right twin is retired with its own evidence, and the other stays pending (failed first on main).
• New: a reworded entry naming only the shared id retires neither twin.
• Mutations killed: restoring the old find (4 red), returning the first twin when ambiguous (1 red), and dropping the exact pass (2 red).
• Suites: renderer 3,035 of 3,035 pass (Node 24); tsc -b is clean.

🤖 Generated with Claude Code

Review disposition (A Codex, B Pi)

Finding	Disposition
A MAJOR: ambiguous twins still settle by cohort inference (the first twin can be retired delivered-inferred)	Declined as an accepted residual and documented in resolveEntryTarget. A's round-2 targeted count: 203 notification entries after same-id twins in 2,413 transcripts; all 203 carry one exact body and zero are reworded. A confirmed the decline
A minor: containment untested (startsWith survived)	Fixed: a wrapped-twin test from the recorded events
B: surviving mutations on unreachable shapes	Accepted: 0 corpus rows with ≥2 complete bodies in one entry
B: phantom "unconfirmed" row for the never-delivered twin	Accepted by design; strictly better than main's false decisions log

Round 2: A MERGE-READY (decline confirmed); B MERGE-READY.