feat(mcp): return from index_codebase without waiting for the first index

[kyle-elliott]

Part 3 of 3 for #37, plus follow-up #46. Stack: #42 → #43 → this PR → #46.

Closes #37

Problem

index_codebase blocked until the whole first index finished embedding. In the reviewed sessions, 4 of 5 calls failed with host timeouts at 30–60 s.

Change

• index_codebase returns right after the codebase is registered. It names the codebase, the path, and sync_status. A first index that is already running reports "already in progress" instead of waiting. A failed first index is still retried.
• FirstIndexPhase (src/mcp/readiness.rs): Registering, Syncing, Embedding { job_id }, Ready, Failed { reason }. It is derived purely from the gate state and the coordinator's last job.
• CoordinatorStatus.first_index (#[serde(default)], so a status body from an older daemon still decodes). sync_status and semctl daemon status show a first index: line.
• Docs, skill, and evals: index_codebase no longer blocks; agents poll sync_status every 10–15 s and use local tools meanwhile. New "When a tool fails" rule: follow the next: line, retry once only when it says the call is retryable, otherwise use local Read/Grep.

Verification

• Full gate passes, including the OMP adapter Bun test.
• Tests: index_codebase returns while the gate is pending; concurrent, completed, and failed gates; the state-to-phase table; old-daemon status decoding; the sync_status phase line; the integration test now expects index_codebase to answer before the job completes.

Known issue (not from this stack)

In a git worktree, the pre-commit hook fails codebase::tests::git_root_preserves_native_bytes_and_trailing_whitespace and nested_git_directory_resolves_to_the_worktree_root, because git init in those tests inherits the hook's git environment. The tests pass outside the hook. All three commits used --no-verify for that reason only.