feat(memory): user work-profile injected at SessionStart

[fazleelahhee]

Summary

Companion to #71 (project_summary). Where project_summary
describes what the project IS, this describes how the user works
on it — cadence, top-touched files, recurring rollup themes,
decision volume — so each new Claude/Codex session opens with a
"you typically work in X area, ship Y sessions a week, here are
the themes that keep coming up" preamble.

Why both PRs together help

Each Claude/Codex session previously had to re-derive:

• what the project is → fixed by feat(memory): project summary + last-3-sessions in SessionStart resume #71 (project_summary)
• how the user typically works in it → this PR (work_profile)

The two are surfaced as separate Markdown blocks in
build_session_resume, written by independent code paths. Either
PR can merge first; the resume just renders whichever blocks exist.

How work_profile is derived

Pure-SQL aggregation over existing memory tables — no LLM, no
embeddings, no network call.

• sessions.started_at_epoch + prompt_count → cadence line
• code_areas GROUP BY file_path → top N most-touched files
• sessions.rollup_summary tokenised, stop-word-stripped, frequency-
  ranked → recurring themes (must repeat ≥ 2× to count, so
  one-off jargon doesn't pollute)
• decisions COUNT → "N decisions on file"

Stop-word set reuses grammar._FILLERS_ULTRA and adds a small
coding-prose-specific extension (fix, add, update, use, …)
so domain terms (cache, retry, auth, parser) survive.

What lands

• memory.db v4: work_profile table (additive migration; safe
  for existing dbs, tolerated when partially-migrated)
• src/context_engine/memory/work_profile.py: builder + upsert +
  load + is_stale + format_profile_block + refresh_work_profile
• build_session_resume prepends the new block; degrades cleanly
  if v4 table is missing
• cce profile CLI (with --force) — shows + refreshes
• cce init runs a baseline refresh after the initial index so
  the very first SessionStart already has data

Tests

21 new cases in tests/memory/test_work_profile.py:

• schema migration
• cadence (skipped <2 sessions; reports span + avg + last-active)
• top_files ranking + empty case
• themes extraction with stop-word strip + min-repeat enforcement
• decision count
• format_profile_block omits empty sections; pluralises "decisions"
• refresh respects TTL but rebuilds on force=True
• build_session_resume includes the block; tolerates missing v4
  table; empty on virgin state

Deferred to a follow-up

Mirror the same block into the MCP context-engine-init bootstrap
prompt so Codex CLI (no hook support) gets the same content via
its system-prompt path. Data + formatter are ready — only wiring is
left, same as the project_summary PR.

Verification

• pytest -n 4 → 896 passed, 1 skipped, 0 failed
• ruff check on every changed file → clean

Test plan

• cce init on a fresh project — passes silently (no
  session history yet, profile block stays empty)
• After 3+ Claude Code sessions, run cce profile — see the
  block populate
• cce profile --force after editing rollups — confirm the
  generated_at_epoch advances
• Restart Claude Code — confirm SessionStart injects the new
  "Your work profile" block at the top of the resume
• Older memory.db (pre-v4) opens cleanly and the resume falls
  back to the old shape