feat(mine): graceful Ctrl-C handling — clean message + exit, no traceback

[igorls]

Problem

mempalace mine <dir> on a large directory is slow (thousands of files × embedding + upsert). Users commonly Ctrl-C if they realize they pointed at the wrong dir, or just want to stop and come back later. Today this produces a bare KeyboardInterrupt traceback through multiple stack frames — something like:

File ".../mempalace/cli.py", line 168, in cmd_mine
    mine(project_dir=args.dir, ...)
File ".../mempalace/miner.py", line 898, in mine
    drawers, room = process_file(...)
File ".../mempalace/miner.py", line 730, in process_file
    added = add_drawer(...)
File ".../mempalace/miner.py", line 662, in add_drawer
    collection.upsert(...)
KeyboardInterrupt

This is ugly, and the user has no signal about:

• How many files were processed before the interrupt
• Whether partial progress is safe (it is — mine is idempotent on re-run via deterministic drawer IDs)
• Whether anything needs cleanup

Proposal

Wrap the main mine() loop in a try / except KeyboardInterrupt block and emit a clean summary:

  Mine interrupted.
    files_processed: N/M
    drawers_filed:   K
    last_file:       <filename>

  Re-run `mempalace mine <dir>` to resume — already-filed drawers are
  upserted idempotently and will not duplicate.

Exit with code 130 (standard for SIGINT) so shell scripts can distinguish interrupt from error.

Edge cases

• Mid-upsert interrupt. ChromaDB upsert() calls are per-batch atomic; a half-processed file is fine because re-running produces the same drawer IDs.
• Mid-closet-build interrupt. Closets are rebuilt per file rather than accumulated mid-file — interrupting between files is clean; interrupting during closet LLM/regex pass for a given file means that file's closets will be recomputed on re-mine (acceptable).
• PID file guard. The existing per-source-directory PID file (fix(hooks): PID file guard prevents stacking mine processes #1023) should be cleaned up in a finally so a future mine doesn't see a stale PID.
• Second Ctrl-C. If the user hits Ctrl-C again during the summary print, fall through to the default handler. Don't try to handle everything.

Tests

• Unit: mine() raises KeyboardInterrupt internally → caller sees exit code 130 and cleanup message
• PID file is removed on interrupt
• Re-running after interrupt does not double-file drawers (already covered by idempotency tests; add one that specifically interrupts-then-resumes)

Out of scope

• Checkpoint-to-disk so re-mine skips work already done (current path already avoids re-embedding via dedup, sufficient for now)
• Making individual file processing interruptible at sub-file granularity