RecordGrounds is an unlocked read-modify-write so a concurrent second grounds write silently discards the first

RecordGrounds is an unlocked read-modify-write so a concurrent second grounds write silently discards the first

Found by the round-2 adversarial security review of build/itd-179.

Two sessions each running abcd intent ready itd-N --grounds "..." both read the same record, each appends onto its own snapshot, and the later write wins. Both processes print the "recorded grounds" receipt and exit clean, so nothing reports the loss: the groundsWriteIsReadable count check compares each writer's own before/after pair, which stays consistent. Reproduced in 40 trials of two concurrent RecordGrounds calls on one record; 39 lost an entry.

This falsifies the append-only guarantee the same file documents at internal/core/intent/grounds.go:50-55 ("Recording is APPEND-ONLY: a second gate decision adds an entry beside the first rather than replacing it").

It also reopens a defect already resolved one day earlier in this same package: iss-2608300235388164, whose body records that stampPlanned was a read-modify-write with no lock, and whose fix put the sequence under withIntentMintLock at internal/core/intent/lifecycle.go:319 with the comment naming exactly this failure. The branch adds a second writer to the record and does not take the lock.

Remedy: wrap the readRepoFile -> appendGroundsBullet -> writeIntentFile sequence in withIntentMintLock, as stampPlanned does.

CORROBORATED independently by the round-2 ruthless review, which graded it FIX-FIRST, reproduced it 20/20 with two goroutines (zero errors returned), and verified the remedy: applying withIntentMintLock around read -> append -> groundsWriteIsReadable -> writeIntentFile in a scratch copy takes the reproduction to 20/20 green with go test and go test -race ./internal/core/intent/ both staying green.

Two adversarial reviewers, in fresh contexts and by different routes, landed on the same defect and the same one-line remedy. That is the strongest signal this cycle has produced for a single finding.

Grounds

• pursued: we expect an append-only contract to hold only where one advisory lock spans the whole read-modify-write, and a concurrent-loss reproduction that stays green after the lock is what would show it wrong