…op to task-bridge (#352)"

This reverts commit 0502234.

First half of the D+C DM-fallback redesign owner approved after the
2026-04-15 DM-flood post-mortem. This PR lands D (separate retention
policy) as a standalone hygiene win. C (event-driven fs.watch wiring
for the DM fallback) will stack on top as a follow-up PR.

### What this does

- Adds `src/archive-stale-results.py` — a small one-shot script that
  walks `results/*.txt`, moves any file whose mtime is older than
  `$RETENTION_HOURS` (default 24) under `results/archive-YYYY-MM-DD/`.
  Files inside existing `archive-*` subdirectories are never touched.
  Honors `DRY_RUN=1` for safe preview.
- Invokes it from `src/startup.sh` right after `mkdir -p tasks results
  data` — i.e. before any long-running service starts iterating
  `results/`. Any accumulated backlog from a prior quiet day gets
  swept out of the service iteration paths at boot time.

### Why this first, before any DM fallback rewiring

Today's incident proved that `results/` accumulates long-dead files
indefinitely: task-*, question-*, briefing-*, insight-*, friction-*
from voice-originated work that was never spoken because voice wasn't
connected at the time. Any code that scans `results/` for delivery —
my reverted #352 task-bridge wiring, #349's discord-bridge polling
loop, or any future design — is going to re-deliver the entire
backlog on first tick after a cold boot unless the directory is
actually kept small.

D (retention) fixes the underlying hygiene concern independently of
the DM-fallback question. Even if we never wire fallback again, the
directory staying bounded is an unambiguous win. With D in place, C
(the event-driven follow-up) can use `fs.watch` without needing a
complex "pre-existing files at boot" escape hatch — D has already
moved them out of the way.

### Retention window rationale

Defaulted to **24 hours**. This is a judgment call:

- Too short (e.g. 1h) — risks archiving same-day results the user
  might still want the voice agent to speak when they reconnect.
- Too long (e.g. 7d) — bounds the accumulation less aggressively.
- 24h feels right: anything older than one full day is effectively
  dead mail. Overridable per-run via `RETENTION_HOURS=N`.

### Verification

- `bash -n src/startup.sh` clean
- `ast.parse` clean on `archive-stale-results.py`
- Dry run against fresh clean results/: "no stale files to archive"
- Fabricated a 2-day-old test file, DRY_RUN=1 correctly flagged it,
  then real run correctly moved it to `results/archive-2026-04-15/`.
- Does not affect `results/calls/` (subdirectory, skipped by
  `f.is_file()` check).

### What's not in this PR

- **C (event-driven fs.watch wiring)** — stacks on this. Will replace
  the current `readdirSync(RESULT_DIR)` scan in task-bridge.ts with
  an fs.watch listener that only reacts to newly-created files.
  Separate PR, coming next.
- **Full architectural redo of DM fallback delivery** — out of scope;
  the pause-on-DM-fallback rule from the post-mortem still holds.
  This PR is pure hygiene.

### References

- Post-mortem: `notes/post-mortem-dm-flood-2026-04-15.md`
- Feedback memory: `feedback_results_dir_is_graveyard.md`
- Owner's D+C decision: Discord 23:20 local

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>