feat: hydrate validate (dry-run coherence findings)

[rennehan]

Summary

Adds hydrate validate — the agent's dry-run gate in the walk → validate → commit loop.

It lowers the current stage into the same typed delta batch commit builds (staging::lower), POSTs it to POST /v1/branches/{id}/validate, and prints the server's coherence report. Unlike commit it is non-mutating: it never applies the batch and, critically, never clears the stage.

Exit code

• 0 when there are no error-severity findings.
• New distinct code 5 (exit::VALIDATION) when there are — so an agent can gate a loop:

  hydrate validate && hydrate commit
  

• The findings always print regardless (human list, or the verbatim {valid, findings[]} with --json); the exit code is only the pass/fail signal. A transport/parse failure keeps its existing code (network 6, generic 1, ...), so "found errors" is never confused with "couldn't reach the service". 5 is unused by the existing set (0/1/4/6).

Backend dependency

Depends on the merged C04 coherence-findings backend contract. POST /v1/branches/{id}/validate now returns {version, project_id, branch, valid, findings[]} with Finding {code, severity, locator, message}, replacing the previous first-failing-delta shape.

This PR re-vendors openapi.json (verbatim from the backend's regenerated spec) and regenerates wire/ via scripts/regen-wire.sh to pick up the new Finding / ValidateResponse schemas. wire/ is generated, not hand-edited; the CI wire-drift check is idempotent against the committed spec.

Changes

• src/cmd/validate.rs — new verb (prepare/render/exit-code split, all unit-tested).
• src/client/mod.rs — validate_deltas(branch_id, body) wrapping the generated call (mirrors apply_deltas).
• src/exit.rs — SUCCESS = 0, VALIDATION = 5.
• src/cli.rs, src/cmd/mod.rs — verb wiring (finish_with_code maps the handler-chosen success code).
• src/cmd/guide.rs, README.md, tests/help.rs — surface/docs/help updates.
• openapi.json, wire/** — re-vendored spec + regenerated client.

Tests

TDD. New coverage: clean report → exit 0 + valid in both modes; error findings → exit 5 (distinct from conflict/network/generic) + full findings printed; --json vs human parity; warning-only findings are advisory (still pass); mixed findings gate on the error; and prepare lowers the stage without clearing it (the non-mutating invariant). cargo test (326 unit + help suite), cargo fmt -p hydrate -- --check, and cargo clippy --workspace --all-targets -- -D warnings all green.