Three independent observations surfaced while smoke-testing v3.3.3 against real Claude Code transcripts. None block the 3.3.3 cut; filing for 3.3.4 triage.

1. _wing_from_transcript_path regex is macOS-centric

File: mempalace/hooks_cli.py:490-505

The regex only matches -Projects-<project> in the transcript path:

match = re.search(r"-Projects-([^/]+?)(?:/|$)", normalized)

Claude Code encodes the full source directory into the project folder name. Users whose code lives outside ~/Projects/ — e.g. ~/dev/, ~/src/, ~/code/, or any Linux convention — get transcript paths like ~/.claude/projects/-home-user-dev-MyProject-myproject/session.jsonl with no -Projects- segment, so they fall through to the wing_sessions fallback and lose the per-project diary scoping that #659 was meant to deliver.

Repro: Stop-hook invocation with transcript_path=~/.claude/projects/-home-<user>-dev-<Parent>-<project>/*.jsonl → diary checkpoint lands in wing_sessions instead of wing_<project>.

Suggested fix: broaden the heuristic to also match the last path segment of the encoded project folder (after the final -), not just the -Projects-<x> slice.

2. mempalace_diary_read doesn't normalize empty-string wing

File: mempalace/mcp_server.py — tool_diary_read handler

#1097 fixed mempalace_search to treat wing="" / room="" as "no filter" (LLM agents frequently default optional string parameters to ""). The same fix isn't applied to mempalace_diary_read — passing wing="" returns {"entries": [], "message": "No diary entries yet."} even when entries exist under named wings.

Repro:

diary_write(agent_name="x", entry="foo", wing="mempalace-dev")  # lands in mempalace-dev
diary_read(agent_name="x", wing="mempalace-dev")                # finds it
diary_read(agent_name="x", wing="")                             # returns empty, should return all

Suggested fix: apply the same empty-string → None normalization from #1097 to diary_read (and audit other tools with optional wing/room kwargs).

3. Stop-hook auto-mine subprocess ignores MEMPALACE_PALACE_PATH env

Observed: When the Stop hook kicks off an auto-mine, the mined drawers land at the default ~/.mempalace/palace regardless of a MEMPALACE_PALACE_PATH override set in the parent env.

Repro: Set MEMPALACE_PALACE_PATH=/tmp/alt-palace in the calling environment, fire the Stop hook (mempalace hook run --hook stop --harness claude-code with a real transcript). The hook's own diary checkpoint respects the env, but the mined drawers from the auto-ingest land at the default path.

Impact: Low for single-palace users (the common case). Can cause drawers to silently disappear into the wrong palace for anyone using multiple palaces or testing with a scratch palace.

Suggested fix: audit the subprocess-launch sites in hooks_cli.py and ensure MEMPALACE_PALACE_PATH / MEMPAL_PALACE_PATH are propagated explicitly (not just inherited).

All three found on develop @ 9947ad0.