[gurinderu]

What this changes

Re-review round detection — finding the previous review run for a branch — was
done by handing a haiku agent a prose recipe: read ~/.craft/runs/index.jsonl,
pick the newest matching line, shell out for ancestry, rebuild a file path from a
timestamp, read that JSON, return selected fields. Meanwhile the same file already
contained a deterministic selectPriorRound that was never called — dead code.

This is the same class of defect the header of lib/craft-log-run.mjs says was
already removed from the write path ("a model was formatting bytes it had no
business formatting"). The read path never got the same treatment.

craft-log-run.mjs grows a prior-round subcommand that owns the whole decision:
tolerant index read, selection via the existing pure selectPriorRound, ancestry
via git, detail-record load, one line of JSON out — and on any failure a clean
"no prior round" with exit 0, so losing it degrades the review instead of aborting
it. The sandbox has no filesystem access, so an agent still transports the result;
it no longer decides anything.

Why it mattered

Measured on the real run store before this change:

• index rows exist with project: "." — a legacy bug where the reviewed repo was
  never resolved to an absolute path. The prose project=pwd match failed them
  silently.
• index.jsonl contains two blocks of pretty-printed multi-line JSON. A strict
  reader chokes; a "best-effort" model reports "no prior round found".

Either way the re-review chain broke without a word and the round restarted from a
blank ledger.

Two review rounds on this branch found five issues, all addressed here:

• a prior round could be taken from an unrelated repository via the "."
  candidate, and the search stopped at the first rejection instead of falling
  through to an older valid round;
• the tests asserted around the two behaviours the change exists for rather than
  through them;
• the output still round-tripped through a strict schema, so one older ledger entry
  missing a key lost the round anyway — the subcommand now normalises every entry
  to the schema shape;
• path.resolve ran in the transport agent's process, against the wrong base;
• the no-"." invariant was enforced at the CLI rather than in the library, so a
  direct call reinstated it. Now enforced where it is stated, with a test that
  fails if the guard is removed.

Fixing the first of those surfaced a false premise: records were not already
stored with an absolute project path (writeRecord kept --project verbatim and
the checkpoint/finalize call sites passed .). Dropping the "." candidate alone
would have broken re-review for the normal case, so project resolution is fixed at
those call sites too.

Scope — what this does NOT close

The primary defect in re-review is finding identity: fingerprint() hashes
model-authored text (file + symbol + ruleId + shingle(title)) and matchesPrior
gates on a 0.6 title-overlap threshold plus a symbol veto. Measured across two
rounds of one branch: 82 ledger entries each, 59 findings in files shared with the
prior round, 1 matching fingerprint, 2 recognised by matchesPrior, and
every entry still disposition: open. That is untouched here and needs its own
change.

Verification

node --test 169 pass / 0 fail; check-workflows 6/6 parse; check-skills and
check-evals clean.