v0.24.0 — Rule 0 ships to adopters

[0.24.0] - 2026-09-04

The Rule 0 instrument now ships. Four files in every generated project pointed adopters at
scripts/framework-check.mjs — a path that exists only inside the create-cmp repo. The tool it
named has never been in an adopter's tree, so the fast way to calibrate a gate was advertised
and withheld, and the expensive way was the only one available.

Added

• qa/framework-check.mjs — Rule 0, in your own tree. Proves the lane returns a fast
  deterministic PASS and FAIL through the real machinery: runner, receipt, Stop hook. No Gradle,
  no device, no network. Measured on a fresh scaffold: 7 plants, 1.9 s.
  • Plants are derived from the tree, not hardcoded — a project with no specs, no flows or
    no Kotlin tests still gets the region plants, and every plant it cannot make is reported
    with its reason. It refuses rather than printing PASS over an empty plant list.
  • Restores everything it touched, including the receipt and the README badge — a smoke
    receipt is refused as done-evidence, so leaving one in place would overwrite a real L1/L2
    receipt with one that proves nothing. Empty directories it created are removed too.
  • Refuses to start when a file it plants into has uncommitted changes.
  • --budget-ms (default 5 s/cycle) reports any calibration cycle slower than Rule 1's stated
    "seconds each" — the bound was prose, and prose does not refuse.
• qa/lib/framework-check.mjs — the plant selection and run assessment as pure functions,
  unit-tested without a scaffold (27 new tests; suite 1306 → 1333).
• scripts/framework-check.mjs proves the shipped twin. A sixth leg runs
  qa/framework-check.mjs inside the scratch app it already stamps and asserts the tree is
  byte-identical afterwards — the tool we hand adopters is proven on the tree as shipped, not
  merely shipped. (Its first version compared git status in a directory that is not a repo and
  therefore compared nothing; it hashes the tree now, build output excluded.)

Added

• qa/lib/agent-hold.mjs + qa/plan.mjs --hold/--beat/--release — liveness for a working
  agent. Two failures reported from payment-blueprint on 2026-09-04 were the same missing
  fact. (a) There was no cheap way to answer "is the agent working or wedged?", so the question
  was answered by filesystem archaeology — twice with broken instruments (a find that excluded
  build/, the only directory a proof run writes to; then a find -newermt that reported zero
  writes in 45 minutes while ls -lT showed one at 10) — and a healthy agent was killed
  mid-proof. (b) The Stop hook fired ~15 times in one evening while the correct action every
  time was to wait, because it cannot tell a receipt stale from laziness from one stale because
  a subagent is mid-commit on a half-adopted port.
  • A hold changes the ADVICE, never the verdict. The hook still refuses — a file an agent
    writes about itself must never be able to end a turn. It follows the precedent already set
    for a lane in flight: same refusal, different instruction.
  • It explains only the two refusals a working agent legitimately causes — no receipt yet,
    and a tree that moved under a PASSing receipt. A FAIL, a forgery, a skipped device tier or
    an unreadable surface are never explained by a hold, because there the hold is not the cause
    and offering it would send the reader to wait on an agent when a test is red.
  • Heartbeat older than 5 min reads as a crashed writer; a hold older than the 45 min ceiling
    reads as a wedge and the alarm returns — a heartbeat proves a process is alive, never that
    it is progressing.
  • Ephemeral like its siblings: gitignored and excluded from the receipt's hashed input surface,
    so a note about who is typing can never invalidate a receipt.
  • 19 tests, including the whole safety asymmetry (suite 1333 → 1352).
• The orchestrator agent now claims, beats and releases the hold, and owes a visible line per
  subagent start and finish.

Fixed

• Signing N artifacts costs one receipt invalidation, not N. A signature changes
  qa/approvals.json's status, and the approvals gate reads status — so a signature
  legitimately moves the receipt's input hash. Signed one at a time after a green lane, that
  means N signatures kill the receipt N times and cost N lane re-runs and N bookkeeping commits.
  Measured in payment-blueprint's log (2026-09-04, 00:12–03:54): 21 commits, of which 7 were
  code, 6 docs and 8 pure bookkeeping ("sign the approval", "bind the green receipt"); every
  code change cost two to three commits and fired the Stop hook after each. It compounds as a
  repo accumulates governed artifacts, which is what makes a session start fast and grind later.
  • node qa/approve.mjs <a> <b> <c> --as "…" signs them in one write — the idiom
    --reopen-feature already established ("one recorded change, not N commands").
  • The CLI states the other half with its reason attached: sign before the final lane run,
    so the receipt you keep already covers the signatures.
  • An unrecognised flag is still refused by name rather than filtered out silently.

Changed

• GATE-RULES Rule 1 names its instrument at the point of authoring. Rule 1 states a
  property — this gate fails, by name, on its own violation — not a procedure to perform by
  hand, and it now says so where a gate author reads it rather than inside a Rule 0 aside.
  Measured cost of the procedural reading, three occurrences in one adoption: a wave briefed to
  plant → ./gradlew → confirm red → revert → build, at 30–60 s per cycle, ~38 minutes with
  nothing committed. It had violated Rule 1's own "seconds each" from the first cycle, and
  nothing noticed, because a bound in prose cannot refuse. A plant kept in the instrument is run
  by everyone forever in milliseconds; a plant performed by hand is run once and lost.
• The orchestrator agent has standing to refuse an expensive instruction. It could already
  escalate before changing a number; it now costs a mandated mechanism the same way —
  instances × seconds-per-instance against the stage budget — names a cheaper instrument if one
  exists, and answers "what in this repo already does this, and why is it insufficient?" before
  briefing any mechanism. It also owes visible progress: a line per subagent start and finish, a
  status upward for anything past ~5 minutes with no output, and proof wall-clock as its own
  report line. A silent agent is indistinguishable from a wedged one.
• The four dangling scripts/framework-check.mjs references in shipped files now name
  qa/framework-check.mjs (verify.mjs, lib/steps-cmp.mjs, lib/evidence-badge.mjs,
  docs/USAGE.md).