[malihashar]

In short

txcript check is a new command that tells you what converting a session to another agent would lose, before you do it. This is step 2 of 3: the library that does the measuring (txcript::fidelity). It adds no command yet, and nothing visible changes for users.

Part 2 of 3 for #92. It builds on #93.

How to review this PR

This branch contains the commit from #93 too, until that PR is merged. Please review only the last commit, feat(fidelity): measure what converting a session into a harness loses. After #93 merges, this PR will show only its own changes.

What this PR adds

A new library module, txcript::fidelity. It answers one question: if I convert this session into harness X, what will I lose?

let report = txcript::fidelity::assess(HarnessId::Codex, &transcript)?;
let all = txcript::fidelity::check_all(&transcript);

A Report lists each difference as lost, changed or synthesized (made up by the target), grouped by what it is:

Severity	Examples
content	text, thinking, tool calls and arguments, results, images, artifacts, error flags
accounting	usage, stop reasons, the model on each message, thinking signatures
metadata	title, model, git branch, working directory

The types are serializable, and nothing in the module does any I/O or needs a new dependency.

How it measures

1. Render the transcript into the target's native text (using dispatch from refactor: move the harness text dispatch out of the wasm module #93).
2. Read that text back.
3. Compare the original with the result.

So it reports exactly what the real codec keeps, and it cannot drift when a codec changes. Some choices that keep it honest:

• It compares a flat stream of blocks, not message by message, because targets split and merge messages.
• JSON that is only re-spaced is not loss. Trailing spaces and empty blocks are ignored. Ids and regenerated timestamps are ignored.
• A stop reason or tool name that only changed form is reported once as "changed", not as one loss plus one addition.
• Items that the target only made up are listed, but they never raise the severity.

The measurement is "codec-level": it does not include what a store adds on top, such as artifact files.

Tests

• 48 unit tests: each category as lost, changed and synthesized; JSON spacing is not loss; split and merged messages are not loss; empty sessions; deterministic output; bounded work on odd input.
• A golden table of what every harness loses on one rich session. It is also published in docs/formats/README.md, and a test fails if the two disagree, so the docs cannot drift. When a codec fix changes a result, the table is updated on purpose.
• A property test: no harness panics on generated sessions, and Simple (the lossless format) reports nothing.
• The full suite passes, and cargo fmt and cargo clippy (0 warnings, with and without default features) are clean.

What you should know

• The golden table will change when other codec PRs merge (for example fix(codec): preserve multimodal artifacts and images across harness conversions #46, fix(codec): preserve slash command (Tool::Command) user prompts across harness conversions #47, fix(codec): unify StopReason parsing and finish mapping across harnesses #48, feat(usage): aggregate token usage and expose in CLI view and across harnesses #50 and fix(common): render tool-result content blocks as text, not escaped JSON or base64 #90). I kept it as one sorted line per harness and category so it is easy to update.
• check_all returns one Result per harness, so one codec failing does not hide the others.

What I could not test

Real agent resumes, OpenCode and Cowork writes (I have no opencode binary or Cowork account here), and the real WASM build.