Department notebook: each department remembers what it learned

[TruongGiang2000]

Stacked on #247. Merge that first; this PR's base will then move to main.

What

Each department now remembers what it learned. company.decisions already was the company's notebook: remember_fact, approval extraction, room lock-in, and the Memory panel. It just didn't know which department a fact belonged to.

• DecisionEntry.dept (nil means company-wide). Identity stays scope|topic, so there is one truth per topic: when Sales re-records the price, it replaces Finance's and takes the tag. The cap goes from 30 to 60 because eight departments now share it (~42 KB of the 1 MiB doc).
• Stamped at the writers, Swift only, with no backend change:
  • chat remember_fact uses the turn's deptKey, the same one sent on the request;
  • approval extraction uses the deliverable's department, which was already resolved for the prompt and then dropped;
  • room lock-in stays company-wide.
• The desk (from Ask a department, get that department: room role prompt + its own desk #247) lists <Dept> has noted:, which is up to 8 of the department's own decisions, newest first. It carries the same "never contradict" instruction as the decisions block, and those entries aren't repeated there. With memory off, neither block is composed.
• Memory panel: each row reads runway · Finance · Codepet.

Spec: docs/superpowers/specs/2026-10-09-department-notebook-design.md · Plan: docs/superpowers/plans/2026-10-09-department-notebook.md

Verified

• New tests, each watched red first:
  • DepartmentNotebookTests: the department is stamped, carried through normalize, and re-tagged when another department re-records; old docs without dept round-trip unchanged; the cap is 60.
  • The chat turn stamps its department, and an ordinary turn leaves it nil.
  • Approval stamps the deliverable's department.
  • The desk shows only its own notes, newest first and capped, never repeats them, and shows nothing with memory off.
• Two fixes came out of a fresh whole-branch review, each with a test that failed first:
  • desk notes had lost the "never contradict" line;
  • excluding notes by identity could drop a second entry on the same topic from the prompt.
• Suites passed: Decisions ×3, ChatContext ×7, CompanyStoreChatTests, CompanyStoreRunTaskTests, ApprovalParityTests, CompanyStoreChatRunTests, VirtualCompanyDecisionTests, MemoryToggleTests. The one host restart is the known 26.2 isolated-deinit crash.
• Not yet checked in the running app. The machine was in use, so I didn't drive it. Steps:
  1. Type "Ask Finance: our runway is 8 months." Expect a Noted chip.
  2. Press + New, then "Ask Finance: how long is our runway?" Expect 8 months.
  3. Open Settings → Memory. The row should read runway · Finance.

Deferred

• The cap of 60 doubles the decisions text sent on every chat and extraction call. This hasn't been measured.
• Small test gaps:
  • the ordinary-turn test doesn't send a department turn first;
  • no test covers a nil re-record clearing the tag;
  • no desk-level test checks that Finance's desk empties once Sales re-records.

🤖 Generated with Claude Code

[TruongGiang2000]

Checked in the running app at fe781d8 on a real account:

1. "Ask Finance: for the record, our runway is 8 months at current burn." → Crash · Finance replied, and the Noted chip read runway — Runway is 8 months at current burn (as of 2026-10-09).
2. + New → "Ask Finance: how long is our runway?" → "Eight months, as of today — that's the one figure Finance has on record." It still said cash and burn are not on record.
3. Settings → Memory: the row reads runway · Finance · codepet. I deleted the test fact afterwards.

Side observation, not from this PR: the reply put the 8-month deadline at "June 2026", which should be ~June 2027. The model seems not to know today's date in Ask.