fix: bug fixes to Rust core

[quangdang46]

Summary

Tier‑1 sync of four small, isolated upstream bug fixes from
milla-jovovich/mempalace (Python) to the Rust port. Each fix mirrors a
specific upstream commit and ships with a regression test.

#	Upstream commit	Issue	Rust file
1	ead2c5d (#1004)	detect_room substring match misroutes (views ⊂ interviews)	crates/core/src/miner.rs
2	40e2c8b (#1156)	Exporter doesn't refuse symlinks at output targets	crates/core/src/exporter.rs
3	0d1c1fb (#925)	Diary skip‑if‑unchanged drops same‑size edits (teh → the)	crates/core/src/diary_ingest.rs
4	2a0ed0c (#1155)	closet_llm bails on first JSON decode error instead of retrying	crates/core/src/closet_llm.rs

Larger upstream gaps (sync.py, source adapters, palace lock, hooks PID
guard, LLM init flags, …) are intentionally out of scope for this PR
— they will land as follow‑ups so each can be reviewed independently.

Per‑fix details

1. miner: token‑boundary detect_room — adds tokens() /
name_matches() helpers that split on -, _, ., / and replace the
raw String::contains checks. Before: every file under
project/interviews/ got routed to a "views" room. After: only true
token matches route. Regression tests:
test_name_matches_token_boundary,
test_detect_room_does_not_misroute_substring,
test_detect_room_matches_separator_bounded_token.

2. exporter: refuse symlinks at output targets — adds
reject_symlink() defense‑in‑depth check. A pre‑placed symlink at the
export path would otherwise redirect markdown writes anywhere the
symlink points (e.g., system directories). Mirrors the miner's
input‑side caution. Tests:
test_reject_symlink_allows_regular_path,
test_reject_symlink_allows_missing_path,
test_reject_symlink_blocks_symlinked_dir (unix‑gated).

3. diary_ingest: hash‑based change detection — switches the skip
gate from byte length to SHA256(content). Same‑size edits (typo fixes,
single‑char swaps) used to be silently dropped, violating the
verbatim‑storage guarantee. Legacy state files without content_hash
keep using the size gate so a post‑upgrade run does not re‑ingest
every untouched diary. Tests:
test_state_entry_legacy_format_deserializes_without_content_hash,
test_state_entry_round_trips_with_content_hash,
test_ingest_diaries_detects_same_size_edit.

4. closet_llm: retry on JSON decode — local LLM runtimes
occasionally produce truncated / partial JSON under load; the previous
bail‑on‑first‑failure path made the existing retry loop dead for that
case. Now JSONDecodeError shares the same exponential‑backoff path as
HTTP 429/503. New RegenerateError::InvalidJson variant. Tests:
test_regenerate_entry_retries_on_invalid_json,
test_regenerate_entry_returns_invalid_json_after_all_retries_fail
(use a small in‑process mock HTTP server).

Local verification

cargo fmt --all -- --check       # clean
cargo clippy --all-targets --all-features -- -D warnings   # clean
cargo test --workspace --all-features                       # 380 passed (369 + 11 new)

Review & Testing Checklist for Human

• Sanity-check the name_matches() token semantics — confirm the
  separator set (-, _, ., /) matches what the upstream
  Python uses and doesn't break any pre‑existing room mapping in
  your real palace.
• On unix, manually create a symlink at the export output_dir and
  confirm export now refuses it instead of writing through the link.
• If you have an existing diary state file under
  ~/.mempalace/state/diary_ingest_*.json, run a no‑op ingest first
  and confirm it does NOT re‑ingest every day (legacy size‑gate
  compatibility).

Notes

• Gap report (Tier 1 vs 2 vs 3 triage) and the patch file are attached
  in the originating Devin session for follow‑up PRs.
• The closet_llm test refactor introduces an internal
  regenerate_entry_with_delay() helper purely to keep the regression
  test fast; the public regenerate_entry() still uses the production
  INITIAL_DELAY_MS = 1000.