fix(replay): resolve Maestro failures and resize progress

[thymikee]

Summary

Resolve Maestro variables at the failing runtime boundary so target labels and runFlow paths are actionable without adding new environment or secret flags.

• Keep Maestro-compatible -e/--env KEY=VALUE; resolved diagnostic identifiers appear in failure output, matching what Maestro attempted.
• Treat text entry as the security boundary: resolved inputText payloads are registered with centralized diagnostics redaction at dispatch and remain hidden from failures. Sensitive recorded input continues to use --record-as.
• Preserve typed error metadata and daemon-owned paths; ordinary env values such as 2, on, or ios cannot corrupt diagnostic timestamps, keys, platform fields, or artifact paths.
• Normalize malformed replay env entries as daemon INVALID_ARGS responses instead of allowing the request handler to throw.
• Bound live-progress cleanup after terminal resize so reflow is cleared without erasing completed output in non-reflowing multiplexers.
• Document the public secrets contract and the intentional native .ad/Maestro diagnostic policy in ADR 0012 and replay guidance.

Validation

• Full pre-push gate passed with unit execution serialized for host stability: lint, typecheck, layering, production exports, metadata, build, Fallow, 4,655 unit tests, and smoke tests.
• Focused router, failure projection, Maestro, divergence, support-matrix, and resize regressions passed.

[github-actions]

Size Report

Metric	Base	Current	Diff
JS raw	1.87 MB	1.87 MB	+213 B
JS gzip	598.7 kB	598.8 kB	+128 B
npm tarball	714.2 kB	714.5 kB	+325 B
npm unpacked	2.50 MB	2.50 MB	+1.7 kB

Startup median (7 runs, lower is better):

Scenario	Base	Current	Diff
CLI --version	26.8 ms	26.1 ms	-0.7 ms
CLI --help	54.2 ms	54.0 ms	-0.2 ms

Top changed chunks:

Chunk	Raw diff	Gzip diff
dist/src/reporting.js	+381 B	+156 B
dist/src/session.js	-328 B	-79 B
dist/src/cli-help.js	+160 B	+51 B

[github-actions]

PR Preview Action v1.8.1
Preview removed because the pull request was closed.
2026-07-28 14:55 UTC

[thymikee]

Not ready at 1b2a98c:

1. Owner-action CI failure: Coverage fails because MAESTRO_COMPAT_LIMITATIONS changed its canonical Environment statement while website/docs/docs/replay-e2e.md split/rephrased it; support-matrix.test.ts requires the contract statement verbatim. Align the website boundary text with the support-matrix contract (or deliberately change contract/test/docs together), then rerun the focused support-matrix test and Coverage.
2. P2 — resize regression test is too weak. progress.test.ts claims to verify every reflowed row is cleared but only asserts that output contains one cursor-up sequence. In the 80→20-column case the previous row occupies four physical rows; an implementation that clears only two would still pass. Assert the exact full clear prefix/count (three cursor-up moves and four erase-line commands), preferably derived from the prior visible width and resized column count.

The replay failure/redaction production path otherwise reviewed clean: identifiers resolve at the runtime boundary, input text is excluded from failure artifacts, injected diagnostic values are registered centrally, and no device-specific validation is required.

[thymikee]

P2 — The new four-row cleanup cap reintroduces stale live-progress rows after a large width decrease. The previous line is truncated to the old terminal width, so a 200-column line resized to 20 columns occupies ten rows in a reflowing terminal; Math.min(MAX_LIVE_PROGRESS_CLEAR_ROWS, ...) clears only four, leaving six stale rows. The new test explicitly pins the incomplete cleanup by expecting only three cursor-ups. Please restore full reflow cleanup and assert nine cursor-ups/ten clears for 200→20, or use a capability strategy that distinguishes non-reflowing terminals without weakening reflowing-terminal behavior. The prior support-matrix and test-strength findings are otherwise resolved, and exact-head checks are green.

[thymikee]

Addressed both items:

1. The canonical support-matrix limitation and website/docs/docs/replay-e2e.md now use the same verbatim failure-diagnostics statement (8f6c1f0ac).
2. The resize regression now derives the four reflowed rows from the initial visible width and resized column count, then asserts the exact clear prefix: four erase-line commands and three cursor-up moves (f121ac0f7).

Validation is green with deterministic unit execution: tooling, Fallow, 522 test files / 4,655 unit tests, and smoke tests.

[thymikee]

P2 remains unresolved at f121ac0. The production code still caps cleanup with Math.min(MAX_LIVE_PROGRESS_CLEAR_ROWS, ...). The new 80→20 test proves only the four-row boundary; it does not cover the failing 200→20 case, and the adjacent test still expects just three cursor-ups there. A reflowing terminal needs nine cursor-ups/ten clears, so six stale rows remain. Please fix the implementation and assert the full 200→20 cleanup rather than strengthening a test that stops at the existing cap. Exact-head checks are otherwise green, with iOS Smoke still running.

[thymikee]

Resolved in 791c86d9d.

• Removed the four-row cap. Reflowing terminals now clear the complete previous width: the 200→20 regression asserts ten erase-line commands and nine cursor-up moves.
• Added an explicit terminal capability for non-reflowing multiplexers. The default reporter detects tmux/GNU Screen sessions and uses single-row cleanup there, avoiding deletion of completed output.
• Added a reporter-level tmux regression in addition to the renderer tests.

Validation is green: tooling, Fallow, 522 test files / 4,657 unit tests, and smoke tests.

[thymikee]

Re-reviewed at 791c86d: clean and ready for human review. The resize defect is resolved—the reflowing path now clears all computed rows, the 200→20 regression asserts ten clears/nine cursor-ups, and tmux/screen use the explicit non-reflowing one-row path with reporter-level coverage. Exact-head checks are green.