Sanitize replayed turns before cross-provider request builds

[TheGreatAxios]

Summary

The vendored inference layer exports a repair pass (transformMessages,
createIDNormalizer in vendor/intx-inference/src/transform.ts) that was
never called from anywhere. Consequences:

• Switching provider/model mid-session replayed a foreign provider's
  thinking-block signature verbatim, which the new provider 400s on. The
  poisoned turn is already persisted, so every subsequent request in that
  session fails identically.
• Output-only block types with no cross-provider wire shape (refusal,
  citation, redacted_thinking, audio, video, code execution) crash the
  Anthropic block builder outright once persisted and replayed.
• Dangling tool_calls only got synthesized error results on the reactor's
  gate-timeout path, not on ordinary cross-provider replay.

src/provider/replay-sanitizer.ts wraps the adapter registry
(withReplaySanitizer) so every resolved adapter's buildRequest first
sanitizes persisted history before building the outgoing request:

1. Drops output-only unmappable block types the builders don't handle
   (redacted_thinking, citation, audio, video,
   code_execution_request, code_execution_result) and converts
   refusal blocks to plain text.
2. Strips opaque provider signatures from foreign-model turns so a
   different provider is never handed back a signature it didn't mint.
3. Delegates to the vendored transformMessages for what it already does
   correctly: dropping thinking blocks from foreign-model turns and
   answering dangling tool_calls with the same synthetic
   tool_result/isError shape the reactor's gate-timeout path uses.

Wired into src/provider/inference-dependencies.ts between
loadAdapterRegistry and createDependencies, so it covers the shared
registry used by the primary agent, sub-agents, and the compaction
summarizer.

vendor/ is untouched — this reuses the exported vendor helpers rather than
duplicating their logic.

Test plan

• bunx prettier --check / bunx eslint on touched files
• bun run typecheck
• bun run build
• bun test ./src ./tests ./evals (5122 pass, 1 pre-existing unrelated
  failure in src/agent/lsp-availability.test.ts reproduces identically
  on main in this environment)
• New regression tests in src/provider/replay-sanitizer.test.ts:
  grok-signed thinking history builds cleanly against both the
  Anthropic and Google paths with no signature/thinking on the wire; a
  persisted refusal block doesn't break the Anthropic builder; a
  dangling tool_call is repaired with a synthetic error result

Fixes CL-6905

[linear-code]

CL-6905