fix: harden dispatch replay and minimal discovery (#194)

[James3014]

Closes the remaining source-side reliability gaps in #194 without changing provider-routing authority.

What changed:

• require agent_start.attemptKey at the public MCP boundary so every consequential dispatch is replay-bound before provider effects;
• expose native grep, glob, and ls in minimal mode while keeping the Codex tool surface unchanged;
• propagate MCP cancellation into Pi discovery tools and impose a bounded 10s discovery timeout with attributable DISCOVERY_CANCELLED / DISCOVERY_TIMEOUT errors;
• update affected server tests so deeper guard semantics remain exercised under the required replay identity;
• add focused Pi cancellation/timeout tests.

Verification on the exact Candidate tree bb4511d762f1b6d674655054a696a8f496ba0b01:

• focused Pi discovery tests: 2/2 PASS;
• focused public MCP schema/minimal-mode tests: 2/2 PASS;
• npm run typecheck: PASS;
• npm run build: PASS;
• git diff --check: PASS;
• full src/server.test.ts: 83/84 PASS; sole failure is the pre-existing P1 Convergence: resolve RECONNECT_REQUIRED session snapshot deadlock after server restart #159 restart/rebind ECONNRESET, reproduced identically on clean main 1f1ec946....

Core physical snapshot:

• source tree a224ca6afb1b51030430206dfda022be4a06adfe
• target tree bb4511d762f1b6d674655054a696a8f496ba0b01
• changed paths exactly four, no deletions, no scope escape.

No provider-specific policy is moved into discovery or the host.