fix(trace): preserve tool outcome certainty

[yearth]

Why

A normalized event currently exposes only isError. When it is false, consumers cannot tell whether the tool definitely succeeded or the source log never recorded a final outcome.

This extracts the adapter/data-layer part of #14 so mindwalk can preserve that distinction without committing to a Trace Health UI or product contract.

What changed

• Add optional outcomeKnown evidence to normalized trace events.
• Preserve Claude Code's explicit is_error: false versus a missing is_error field.
• Parse Codex command output into known success, known failure, or unknown.
• Treat structured Codex patch_apply_end.success as a known outcome.
• Keep the trace schema and TypeScript contract mirror in sync.

The resulting states are:

outcomeKnown	isError	Meaning
true	false	known success
true	true	known failure
absent / false	false	outcome unknown

outcomeKnown is optional, so traces produced before this change remain valid under the updated schema. Existing isError: true events still identify failures without the new field.

Scope

This PR intentionally does not include session-health derivation, a health API or CLI command, or any UI. It only preserves source evidence in the shared trace model for future consumers such as judge input, comparisons, or metric annotations.

Validation

• go test -count=1 ./...
• go vet ./...
• make test
• mindwalk trace testdata/claude-session.jsonl smoke check confirmed outcomeKnown: true in exported events

Related to #14.