Build a local failure-to-regression workflow with agent-replay integration

## Priority

P2 — focused reliability or product improvement

## Context

Portfolio audit follow-up (2026-09-08). The most useful product experiment is importing a real coding-agent failure, producing a small replay, and making a regression executable in CI. This is a scoped product proposal; adoption and willingness to pay have not been established.

## Starting points / evidence

Depends on semantic normalization and outcome assertion work. Cross-repository companion: https://github.com/JesseRWeigel/agent-replay .

## Implementation scope

Support one documented transcript format first. Add an import/inspect/redact/export flow and emit a deterministic regression fixture using semantic argument and outcome assertions. Coordinate a versioned interchange format with JesseRWeigel/agent-replay. Provide one runnable CLI-to-CI example; avoid building a hosted platform for this milestone.

## Acceptance criteria

- [ ] A synthetic failed run can be imported, inspected, redacted and exported using documented commands.
- [ ] The generated check fails on the bad run and passes on a corrected fixture.
- [ ] Secrets/private paths can be removed before sharing, with a preview of the exported content.
- [ ] Document format/version limits and actionable errors for unsupported input.
- [ ] Provide an external-developer pilot script and success criteria: unaided import, a real caught regression and repeat use; record observations only after actual consented trials.

Keep the change focused on this issue. Use synthetic fixtures for sensitive data and mocked providers for automated tests; do not require paid model calls to run the regression suite.

Start by reading the existing semantic normalization and outcome assertion work, then review the documented transcript format and the agent-replay companion. Define the import, inspect, redact and export flow around synthetic fixtures and mocked providers. Done means a documented CLI-to-CI example imports and previews a redacted transcript, fails on the bad run, passes on the corrected fixture, and states format limits and pilot criteria.

Written by the indexing model from the issue text.