[thangvmt]

Addresses @khenzarr's measurement on flop-labs#97: the warning pass verifies every record three times.

foldTranscript walks the records three times and calls verifyTranscriptRecord in each pass — once to build the per-room map for gap analysis, once inside the hasDeadlineFrame probe, and once in the fold itself. tryDecodeFrame runs twice for the same reason. Ed25519 verification dominates the cost of a large retained export, so the warning feature roughly triples the price of a fold even for lines that never decode as TCLK frames.

This patch builds on the verified-row gap analysis from flop-labs#94/flop-labs#97. Composing that pass with deadline detection and folding repeated signature verification, so this refactor caches verification and decoding once per record and shares the result across all three consumers.

Measurement

Same machine, same input, timing one foldTranscript call after building 2000 deterministic signed records. The unpatched number comes from stashing only src/transcript.ts and rebuilding, so both rows are measured under identical conditions:

before   6792.3 ms / 6788.9 ms
after    2257.7 ms / 2259.2 ms

That is 3.01x, consistent with the 2.93x reported on the issue.

The test pins structure, not time

tests/transcript-cost.test.ts asserts one verification per record rather than a duration, because a timing assertion is fragile on shared CI.

ed25519 is a frozen object, so vi.spyOn fails with Cannot redefine property: verify. The module is mocked through vi.hoisted in an isolated file, and the real implementation is called behind a counter. Keeping it in its own file means that if the mock ever breaks it does not take the existing suites with it, and no injection seam is added to production code for the sake of a test.

The test REDs on the unpatched base with expected 12 to be 5. The 12 is itself informative: the hasDeadlineFrame probe uses records.some, so it short-circuits at the second record rather than walking all five.

Behavior

No protocol or state-machine behavior change; one warning index is corrected for aliased duplicate objects. Carrying the original index on each checked entry also removes the records.indexOf scan inside the gap loop, which was O(N²) on the number of verified rows.

Gates

Per AGENTS.md, recursive with the workspace root:

pnpm -r --include-workspace-root build
pnpm -r --include-workspace-root test

root 107/107, mcp 40/40, worker 33/33 — 180/180.

One conflict to flag

My other open patch against this branch, #1, appends the derived-room firstSeq anchor to the same for (const [room, list] of byRoom) loop. After this refactor the loop iterates checked entries rather than raw records, so that block needs one adjustment when both land:

-      const firstSeq = list.reduce((min, r) => (r.seq < min ? r.seq : min), list[0]!.seq);
+      const firstSeq = list.reduce((min, r) => (r.record.seq < min ? r.record.seq : min), list[0]!.record.seq);

Merge in either order and I will rebase whichever lands second — say which you prefer and I will send it.