[santoshkumarradha]

What

A plan node that ran out of its token budget was released back to pending and re-dispatched with the byte-identical grant and brief, up to MaxOverrunRounds, and bought the same truncated ending each time (observed twice in one run, ~12 USD).

• cmd/codeaf/subharness.go: regrantAfterRunningOut grows the token grant by three halves per re-dispatch, never above overrunGrantCeiling (four flat leaf grants; a fan-in that measured more keeps every token it measured). Wired at the dispatch grant in chat.go; the leaf's wall follows the larger grant.
• internal/resident/runner.go: a re-dispatch that banks nothing beyond what the hand-on release journaled (store.ReleasedTurnsFor) is failed with the refusal named (outOfRoomRefusedFailure) instead of sent round again.
• internal/store/lifecycle.go: ReleasedTurnsFor, NodeRedispatches read off the journal.
• --json envelope: redispatches (absent when none), counted per node of the run's own session; manual row added.
• Tests: TestARedispatchIsGrantedMoreRoomThanTheAttemptThatRanOut, TestAnExhaustedRedispatchThatBankedNothingNewFails, TestAnErrandCountsTheRedispatchesItsOwnNodesBought, envelope key test extended. Change note in docs/changes/unreleased/.

Verification

Spark job 20260917-094736-000551 on e689581ea, go test -count=1 ./cmd/codeaf/ ./internal/resident/... ./internal/store/...:

ok  cmd/codeaf          69.348s
ok  internal/resident   27.160s
ok  internal/store      23.067s

The only commit after that head renames the change note to this PR's number. Not merged; awaiting the owner's look.

🤖 Generated with Claude Code

[santoshkumarradha]

Hi @AbirAbbas, please review as high priority.

[santoshkumarradha]

Also internal/manual on head 85aae36c2: ok (Spark job 20260917-114912-000563), for the envelope row the manual gained.

[santoshkumarradha]

Retargeted to dev at Santosh's request: the branch was rebased from santos/dev onto origin/dev (09299019b), same diff, and builds clean. Spark make pr-ready on this head is fleet job 20260917-130548-000567; result follows.

[santoshkumarradha]

Correction: the Spark job for this head is 20260917-132058-000568 (cpu lane); 000567 was queued behind a stuck gpu-lane job and dropped.

[santoshkumarradha]

Spark make pr-ready on head eafafb160 against origin/dev 09299019b: exit 0 (light gate + touched-package suites, fleet job 20260917-132058-000568, finished 13:37Z).

[santoshkumarradha]

Folded into #1093 (merged into feat/auto-crew at 5e7f0f0); this branch is kept for its history.