fix(subagents): support Pi tools, target workspaces, and Cursor execution

[tt-a1i]

Workflow children could fail before useful work for several independent reasons: Cursor rejected every tool invocation, investigator roles excluded shell/network tools, Workflow had no explicit child cwd, and a local 45-second output timer cut off otherwise live Pi requests. This change makes ordinary and Workflow children usable through Pi's own tools and lifecycle.

• Bridge Cursor's advertised MCP calls into normal Pi tool execution and paired result continuation; native Cursor execution and unknown identities remain rejected.
• Inherit active parent child-eligible tools for built-in roles. Preserve custom allowlists, parent-only exclusions and Plan Mode restrictions.
• Add Workflow working_dir, including target trust, replay/operator identity and worktree cleanup. Inherited roles execute for real on resume; bounded custom read-only replay stays restricted.
• Remove extra child output watchdogs; retain native provider timeouts/retries and bounded cancellation/disposal.

Existing custom role files are preserved: an old explicit explorer.md allowlist still excludes shell/network until deliberately edited. This is an intended widening of built-in role capabilities in normal mode, not a filesystem sandbox. The current installed runtime is not changed by this PR.

Validation: bun run check passed. The full repository-discovered test set passed with Node file concurrency 2: 1435 passed, 1 platform skip; Vitest 30/30. The default bun run test run hit the pre-existing Setup subprocess 8-second startup deadline under local host load; its failure is retained rather than changing that unrelated test. Two independent reviews found no actionable defects. Live Cursor Grok accepted the Pi tool/result round trip; a real Workflow child read outside cwd and fetched HTTPS through Bash; eight concurrent production-runner children passed 8/8 in about 12 seconds. These small acceptance runs do not establish long-audit quality or account rate ceilings.

Investigation and evidence boundaries: research record. Dashboard lag was fixed separately in #422. The independent Windows/headless shell configuration repair remains in #423.

Fixes #424

Final CI on c3809e1f5c39d736054694724dce69a1099fe49d: Node 22.19.0, Node 24, and Windows all passed, including the standard repository commands and package smoke checks. Fixed review base: a7455cd378ef7befa9c7cf099c1fa1e4ee5dc3e3. CI receipt.