…remember

The memory compression path summarises conversation events and can therefore
drop exact user text — names, numbers, sentences the user flagged with an
explicit "remember this" imperative were getting paraphrased or summarised
into generic bullets in `## Key Decisions`, then losing specificity further
on each update cycle. Result: durable memory was missing content the user
explicitly asked to preserve.

This is handled entirely in prompts — no language-specific keyword table,
no regex gate. The compressor LLM already reads the user's messages; we
just tell it to recognise the INTENT (cross-language: 记住/记得/记下/牢记,
remember/keep in mind/note this, 覚えて, 기억해줘, recuerda, запомни, etc.)
and route that content into a dedicated section.

- `summary-compressor.ts buildCompressionPrompt`: new `## User-Pinned Notes`
  section in the structured template. Both the fresh-compression branch and
  the update-existing-summary branch carry an explicit VERBATIM PRESERVATION
  RULE telling the LLM: copy pinned content word-for-word, never paraphrase,
  translate, or truncate; carry it forward unchanged across every update.

- `materialization-coordinator.ts extractDurableSignalsFromSummary`:
  refactored to use a shared `extractSummarySection` helper (escaped-title
  regex) so any section can be pulled out. Now also scans
  `## User-Pinned Notes` and pushes each non-empty line into the
  `preferences` bucket of DurableSignals. Effect: pinned content is
  automatically promoted to `durable_memory_candidate` instead of living
  only in the short-lived recent_summary, where it could eventually age
  out on rotation.

Existing summary-compressor + materialization tests remain green (25/25).
The compressor tests don't assert on prompt text, so the change ships
without touching test expectations.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>