fix(record): the live position is derived, so no PR writes NOW.md (#374)

[localai-bot]

Issue #374. Row ENG-NOW-DERIVED. Spec .agents/specs/now-derived.md. Follow-up to #364.

No product source is touched.

The problem you can't fix by asking people nicely

#364 removed NOW.md's byte budget — the forced eviction — but not the forced write. One line kept every PR in the file:

REQUIRED = {"lifecycle": (STATUS, BENCHMARKS, NOW), ...}

The gate marched every lifecycle move into one shared file. NOW.md conflicted in 5 of the 16 conflicting open PRs at d928e2c3. And d72ac5a3 — "keep the ROCm live-claim row inside the NOW.md budget" — is an unrelated author paying that tax mid-flight, while #364 was removing the constant demanding it.

By the invariant #364 added to AGENTS.md, NOW.md was a documented violation of the rule it shipped alongside.

Relocate the obligation, don't delete it

"The live position must be current" survives — it's just paid somewhere that can hold it.

• A row's next step → the row's own spec, under ## Now. One writer. .agents/specs/ took zero conflicts across the whole Shared record surfaces are a lock: 16/29 open PRs conflict, 13 of them in bookkeeping only #364 sample.
• check-doc-checkpoint.py requires that spec, not the digest — and fails closed, naming the row, when a matrix line links no spec at all.
• scripts/now.py renders on demand from the matrices, .agents/claims/, each spec's ## Now, and open PRs. 105 rows in ~50 ms with no network, degrading to REMOTE_UNVERIFIED rather than to a silent empty roster — "no claims" and "could not ask" look identical in a rendered table, and that's the trap.
• .agents/NOW.md keeps only the authored half and gains a regrowth guard, because removing the table once isn't enough when the decay path is someone re-adding "just one row".

Net: a row-advancing PR writes its own spec and its own claim file, and touches no shared surface.

Two consumers the issue hadn't counted

Both relocated the same way, not deleted:

1. check-release-binary-contract.py pinned a literal | Release | … | row inside NOW.md. Now pinned in release-binary-matrix.md's own ## Now.
2. Two of Shared record surfaces are a lock: 16/29 open PRs conflict, 13 of them in bookkeeping only #364's own tests asserted that adding a row to NOW.md is free. That's now the wrong question — retargeted to authored content, which is what the surviving obligation is actually about.

AGENTS.md

The §Records paragraph ended with a dated measurement naming SHAs — a log entry in the one file this protocol says history doesn't go, and the same accumulating-justification pattern #364 removed from STATUS_RATCHET. Compressed to the rule, plus a corollary naming the real mechanism: when a checker requires every change to touch a shared file, the gate is the defect, not the people tripping it.

Dogfooding

This row's own claim is .agents/claims/CLAIM-ENG-NOW-DERIVED.md — the first real use of #364's one-file-per-claim shape, not a row appended to a shared table.

Evidence

preflight rc=0. Every suite carries the mutation that disarms its own guard:

Suite	Cases	Mutation proven
test_doc_checkpoint	26	restoring NOW to the triple re-breaks the NOW-free case
test_now_render	11	three degrade paths; digest reads nothing from the authored file
test_check_now_current	15	disarming ROW_TABLE_LINE lets the regrown table pass
test_check_gate_commands	29	dropping the baseline entry breaks the exact pin
test_check_release_binary_contract	30	each relocated anchor fail-closed

🤖 Generated with Claude Code