distill

Add any information that is deemed important for a future agent or human engineer working on this project to doc/loom/knowledge, focusing on architectural insights, conventions, and mistakes made (and their resolution if available).

Start from the grouped worklist. Run loom memory pending --group and work the groups in order: corrections (text starting stale-knowledge:, target printed as <file>#<heading>), then mistakes (mistake:), decisions, other. When a mistake is a recurrence of one already in the tree, record a proposal for a hook or a loom plan verify check in concerns.md instead of another paragraph, and resolve the memory as merged.

Corrections first. Apply EVERY corrections-group entry. Correct files IN PLACE with loom knowledge replace-section <file> "<heading>" "<corrected body>" (body WITHOUT its ## heading line) — never loom knowledge update for an existing claim, which only adds to a file rather than fixing what is already there.

Current truth only, outside the mistakes files. Every knowledge file except mistakes.md and mistakes/* states how the tree works now. Rewrite any dated heading or "was", "used to", "supersedes" narrative you find there into current-state text with replace-section; when the history carries a lesson, move it to a mistakes topic and link to it.

Route each new insight to the right file:

• architecture.md — component relationships, data flow, module/dependency graph
• patterns.md — reusable architectural patterns found in the codebase
• conventions.md — naming, structure, and coding standards
• mistakes.md — what went wrong, the misleading signal, root cause, prevention rule, and fix
• concerns.md — tech debt, warnings, known issues
• stack.md / entry-points.md — dependencies/tooling and key files to read first

If doc/loom/knowledge/INDEX.md exists, the layout is hierarchical: each file above is a tier-1 summary, and per-category directories (architecture/, mistakes/, ...) hold tier-2 topic files for deeper detail. Route by size — a finding that fits in roughly 40 lines or fewer goes inline into the tier-1 file above; anything larger goes to loom knowledge update <category>/<slug>, leaving a 2-4 line summary plus a link in the tier-1 file. INDEX.md regenerates automatically on every knowledge write, so there is nothing to run at the end.

Consumers of what you write here do not read these files end to end: each stage session gets a per-stage Knowledge Brief that loom knowledge context assembles by pulling in whichever sections retrieval judged relevant to that stage's task. Write to be retrieved in slices, not read start to finish — prefer one self-contained ## section per finding, naming the concrete symbols and paths inline, because a section is the unit retrieval selects. A finding scattered across sections, or missing its own heading, may never surface to the agent that needs it.

Give every memory taken from an insight a receipt. After the loom knowledge update/replace-section call that used a Note/Decision/Question entry, run loom memory resolve <id> --outcome promoted --target <file#heading>; use merged when it folded into an existing section, discarded --reason "..." when it was a duplicate or wrong, deferred --reason "..." when it needs evidence not available now. Finish with loom memory pending --strict and resolve whatever it lists — this also runs standalone, any time, not only at plan end.

Only record what is non-obvious and durable. Do NOT record ephemeral task details, procedural steps, or anything already captured in the code, git history, or these files. Convert any relative dates to absolute.