Add marker-backed comments and maintainer command protocol

[romgenie]

Local source: coven-github/issues/08-add-marker-backed-comments-and-maintainer-command-protocol.md

Summary

Bot status comments should be marker-backed and edited in place, and maintainer commands should be parsed through an explicit command protocol. The current worker posts start/progress comments, and mention matching is simple substring detection.

Current Evidence

• crates/worker/src/lib.rs posts a starting comment for issue tasks.
• DESIGN.md cites ClawSweeper as the reference pattern for conservative GitHub automation, including durable state, marker-backed comments edited in place, maintainer steering commands, and deterministic gates before repair or merge work.
• ROADMAP.md says to borrow ClawSweeper's edited status comments, maintainer steering commands, and live-state rechecks before mutation.
• crates/webhook/src/routes.rs detects mentions by checking whether the comment body contains @{bot_username} after trimming [bot].

Problem

Repeated bot comments clutter issues and PRs. Simple mention matching is not enough to separate casual mentions from commands with consequences. A hosted coding agent needs a stable command grammar and a single mutable status surface per task or target.

Impact

• Issues and PRs can accumulate duplicate status comments.
• Bot behavior can be triggered by ambiguous or accidental mentions.
• Maintainers cannot reliably steer, pause, cancel, deepen, retry, or approve memory/write behavior.
• Cave and GitHub comments can diverge as status sources.

Proposed Design

Implement marker-backed comments:

• A hidden HTML marker identifies the task, target, and bot/familiar.
• The worker searches for an existing marker before creating a new comment.
• Status transitions edit the existing comment.
• Completed tasks summarize outcome and link to Check Run, PR, or Cave session.

Implement explicit commands:

• @familiar review
• @familiar fix
• @familiar deepen
• @familiar retry
• @familiar cancel
• @familiar remember
• @familiar forget
• @familiar status

Commands should be parsed into typed task intents. Unknown or ambiguous commands should receive a safe clarification response rather than launching work.

Acceptance Criteria

• Re-running the same task updates an existing marker-backed comment instead of creating duplicates.
• Mention parsing recognizes explicit commands and ignores plain text that is not a command.
• Self-comments continue to be ignored.
• Maintainer permission checks gate commands that mutate code, memory, or publication.
• Tests cover duplicate status updates, command parsing, unknown commands, and self-comment loops.
• Documentation lists supported commands and their effects.

Test Notes

Add parser tests independent of GitHub transport. Add mocked GitHub API tests for finding/updating marker comments. Add route tests for ambiguous mention text that should not create a task.