Feature: allow post_tool hook to transform tool output (not gate-only) #1241

Motivation

docs/HOOKS.md documents pre_tool as a gate (exit 0/2). post_tool currently fires for telemetry but cannot modify the tool result returned to the model. This blocks a large class of context-hygiene automations:

• Auto-invoking mcp__headroom__headroom_compress on any tool output above a byte/line threshold.
• Redacting secrets that leaked into command output.
• Truncating or summarizing verbose logs before they enter context.

Today the AGENTS.md "compress outputs > 2 KB" rule is prose-only. Users cannot enforce it mechanically, so per-session compression coverage is near zero even when the Headroom MCP is configured.

Proposed contract

• post_tool hook may print a JSON envelope on stdout, e.g. {"replace_output": "..."} or {"replace_output_ref": "hr_<hash>"}.
• Exit 0 with no stdout = pass-through (current behavior).
• Exit non-zero = fail-open with warning (matches current fail-open semantics).
• Add post_tool_timeout_ms mirroring pre_tool_timeout_ms.

Alternatives considered

• Doing it in pre_tool — impossible, output does not exist yet.
• Client-side hook in the harness — jcode is the harness.
• MCP proxy layer — would need to intercept every tool call type, not just MCP.

Related: ~/.jcode/preferred-tools.md already tells the model to compress big outputs, but nothing enforces it. Version: v0.84.0.