spec: falsifiable-by-replay evidence — ship the probe recipe inside the stamp

[HMAKT99]

Raised by Mike Czerwinski on the launch post.

The insight: a signature proves who said it; a replay proves it could have been true. AKF stamps today are signer==subject (the agent attests to its own work) — an accountability trail, not verification. What makes an attestation reusable by strangers is falsifiability: anyone can re-run the probe and diff against the claimed result.

Proposed: extend Evidence with an optional replay recipe:

• replay.command (e.g. pytest -q), replay.cwd, replay.expected_exit, optional replay.output_digest
• New akf verify <file>: re-runs the recipe and reports CONFIRMED / REFUTED / UNREPLAYABLE — turning claimed evidence into checkable evidence
• akf check could weight replayable evidence above bare receipts — which also gives trust: evidence weighting can be gamed by weak tests (stamp inflation) #125 (stamp inflation) a principled answer: a weak test can still be replayed and seen to be weak

Touches: python/akf/models.py (Evidence), stamp.py, check.py, new verify command.

[HMAKT99]

Design update from the launch-post thread (credit: Mike Czerwinski, follow-up comment):

The seam between this and #129: a replay proves the command produces the claimed digest against whatever inputs are present at replay time. If a dependency/source moved and the staleness flip hasn't fired (or nobody re-ran check since), replay returns CONFIRMED against drifted inputs — the attestation reads freshly re-validated when it actually validated the wrong world. A replayable claim can be provably reproducible and still be reproducibly wrong.

Adopted into the design:

• The replay object carries its own input fingerprint pinned at issuance: replay.input_hash — a digest over the claim's input closure (the stamped file + recorded deps from check: staleness misses transitive invalidation (dependency closure) #124/feat(check): dependency-aware staleness #126 + cited sources from feat: content-hash-at-timestamp attestation over cited sources #129 when present)
• akf verify then splits CONFIRMED into two states:
  • CONFIRMED — replay succeeded AND current inputs match the issuance fingerprint
  • CONFIRMED_DRIFTED — replay succeeded but against inputs that have diverged since issuance (name TBD)
  • plus REFUTED / UNREPLAYABLE as before
• This makes spec: falsifiable-by-replay evidence — ship the probe recipe inside the stamp #128 depend on the input-closure hashing groundwork rather than shipping the two separately and inheriting staleness's blind spot.