fix(palace_graph): skip None metadata in build_graph

[jphein]

Problem

palace_graph.build_graph at line 95 calls meta.get("room", "") unconditionally on every entry in batch["metadatas"]. ChromaDB returns None for drawers without metadata (legacy data, partial writes — same root cause as the entries fixed in #999 and #1094 in different files), and a single None drawer crashes the entire graph build with AttributeError: 'NoneType' object has no attribute 'get'.

This takes out every consumer of build_graph: graph_stats, find_tunnels, traverse, and any palace-daemon-style HTTP endpoint that calls them. Caught 2026-04-25 by a smoke-test script against a 151K-drawer production palace — /stats was 500-ing on a single bad drawer.

Fix

Skip None entries with a continue early-return inside the loop. The graph build is recoverable: build_graph already filters room and room != "general" and wing immediately downstream, so a missing-metadata drawer was never going to participate in the result anyway. The skip is silent because logging every legacy-data drawer would be log spam on palaces that have many of them.

for meta in batch["metadatas"]:
    if meta is None:
        continue
    room = meta.get("room", "")
    ...

Closes the same gap as

• fix(searcher): guard against None metadata in CLI print path #999 (None-metadata audit in searcher.py / mcp_server.py / miner.status, merged 2026-04-18)
• refactor(backends/chroma): coerce None metadatas to {} at backend boundary (closes #1020) #1094 (coerce None metadatas to {} at ChromaCollection boundary, MERGEABLE)

palace_graph.py is a separate read path that the #999 audit didn't reach, and #1094 (when it merges) handles this at the backend layer for new writes — but legacy drawers already in the palace need this defensive check.

Tests

TestBuildGraph::test_none_metadata_does_not_crash constructs a fixture with a None entry mixed between two real drawers and asserts the build completes with the two real drawers processed normally.

🤖 Generated with Claude Code

[Copilot]

Pull request overview

Fixes a production crash in palace_graph.build_graph when ChromaDB returns None entries in metadatas (legacy/partial-write drawers), allowing graph-based endpoints like /stats to remain available even with bad rows.

Changes:

• Skip None metadata entries during build_graph iteration to avoid AttributeError.
• Add a regression test ensuring build_graph completes and counts valid drawers correctly when metadatas contains None.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated no comments.

File	Description
mempalace/palace_graph.py	Adds a defensive None check while iterating drawer metadatas to prevent crashes.
tests/test_palace_graph.py	Adds regression coverage for mixed {dict, None, dict} metadata batches.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[bensig]

Approve. Same shape as #1198 — defensive skip of `None` with a real reproduction case (palace-daemon's verify-routes.sh against the canonical 151K palace, 2026-04-25).

What landed

`build_graph` now skips `if meta is None: continue` rather than crashing on `AttributeError`. The comment is the right shape — names the symptom (`/stats` 500-ing, taking out every consumer of `build_graph`), points at the upstream issue family (#1020 / #999 / #1094), and dates the catch.

Tests

`test_none_metadata_does_not_crash` exercises a 3-drawer collection where the middle drawer has `None` metadata and asserts both real drawers still get processed. Specifically asserts `nodes["auth"]["count"] == 2` so a regression that silently drops both real drawers along with the None one would catch. `pytest tests/test_palace_graph.py` — 24 passed on this branch.

Observation, not actionable

This is the third `None` defense in three review sessions (this one + #1198 BM25 tokenize + the diary_read `wing=""` from #1147). The pattern keeps recurring because Chroma's columnar return shape doesn't carry typed nulls, and downstream code keeps assuming non-None where the data layer doesn't promise it. Worth a follow-up at some point: a `@dataclass` adapter at the `ChromaCollection` boundary that fills None metadata with `{}` and None documents with `""` so consumers don't each need their own defensive layer. Out of scope for this PR.

Ship it.

[jphein]

Thanks @bensig — ready for merge whenever convenient. Tests stable on the fork's daily-driver palace + smoke pass on the develop branch.