ara check --fix: relax value-recovering guard to work on artifacts with pre-existing errors

[EYH0602]

Problem

ara check --fix (added in #41) applies a fix only when the whole artifact re-parses fully clean. On the messy real corpus — where, per docs/ara-format-feedback.md item 13, "none parse cleanly; about half emit errors" — this means --fix usually applies nothing, even when the fixable drift is itself the cause of the errors.

The guard's Ok-gate is here:

// crates/ara-core/src/fix.rs — guard_ara004
let Ok((mc, _)) = cand else {
    return false;   // any remaining error in the edited artifact → reject the fix
};

The guard_alias path (ARA002/ARA003) is similarly conservative: it requires a clean base parse to compute its targeted one-field diff, so an unrelated pre-existing error (duplicate id, cycle, etc.) discards those fixes too.

Evidence (from an ara check sweep over the 14 vendored fixtures)

Detection is correct (0 panics; 2 artifacts flagged with real ARA004 dash-header drift), but --fix applied 0 fixes to both fixable artifacts:

• corpus/speedrun/nanogpt-speedrun — all 10 claims C01–C10 use em-dash headers (## C01 — …), so none parse, producing 29 evidence references unknown claim errors. The dash headers ARE the cause of those errors. Fixing them would recover the claims and clear the errors — but the fixpoint loop applies one fix at a time, and each single fix still leaves the artifact erroring, so every candidate is rejected and nothing moves.
• corpus/rebench/rebench-restricted_mlm — the schema_version: "ara-2.0" shape (no tree:/root:); the base parse errors fatally, so all fixes are blocked. (This one is out of scope — a genuinely different format — and is fine to leave blocked.)

Proposed change

Relax the value-recovering guards (ARA004, and ARA002/003) from "must re-parse fully clean" to "must not introduce new errors" — i.e. accept the edit when errors_subset(cand, base) holds (the edited artifact's error set is a subset of the pre-fix error set) plus the existing targeted-recovery + idempotence checks. This keeps the safety guarantees (never adds an error, never corrupts a file, still idempotent) while letting a fix proceed on an already-imperfect artifact.

Note: errors_subset(cand, base) is already computed in guard_ara004 but is currently dead code behind the Ok early-return. The main implementation wrinkle is that parse_sources discards the manifest on any error, so recovery verification must use the tree-isolated claims_only(...) parse the guard already uses for the base claim set — extend that to the candidate side so the post-fix claim set is readable even when the tree still errors.

Acceptance

• ara check --fix corpus/speedrun/nanogpt-speedrun (on a copy) recovers the C01–C10 headers and clears the resulting unknown claim errors.
• No fix is ever applied that introduces a new error diagnostic.
• --fix remains idempotent; a discarded fix still leaves the file byte-identical.
• Add a corpus-style test with a pre-existing unrelated error to lock in the relaxed-but-safe behavior.

Follow-up from #41. See also the "conservative on an already-broken artifact" concern noted during that implementation.