Bug: harvest can never produce 'context' memories (type missing from valid set)

[raphasouthall]

Summary

`context` is a supported memory type (has a TTL-168h branch in `harvest.py:753`) but the harvester's LLM classifier rejects it. All 13 context memories in the current vault are hand-written via `vault_remember`. This partially explains the skewed type distribution: 38% observation vs 2% context.

Root cause

`src/neurostack/harvest.py:647` — `_llm_classify` valid set:

```python
{"bug", "decision", "convention", "learning", "observation"}
```

`context` is missing. The LLM prompt at `harvest.py:600-604` also omits it. So even if the LLM tried to emit `context`, the classifier would reject it.

Downstream: the TTL-168h branch at `harvest.py:753` is unreachable from harvest — dead code until this is fixed.

Proposed fix

1. Add `"context"` to the valid set at `harvest.py:647`.
2. Extend the prompt at `harvest.py:600-604` to define context: ephemeral state, credentials, URLs, or session-scoped facts with a short TTL.
3. While there: the observation prefilter at `harvest.py:488-492` over-captures (matches generic terms: credential, api-key, endpoint, host, port, url, stored at). Consider routing `credential`-pattern matches to `context` (short TTL) rather than permanent `observation`.

Expected effect

Unblocks the TTL-168h path. Starts populating the context bucket from sessions. Reduces observation noise by routing ephemeral facts to their intended bucket.

Key files

• `src/neurostack/harvest.py:488-493, 600-605, 647, 753`

[raphasouthall]

Acceptance criteria

• harvest.py:647 valid set includes `context`
• Harvest LLM prompt (lines 600-605) defines when to emit `context` (ephemeral state, credentials, URLs, session-scoped facts)
• Integration test: harvest a synthetic session containing "credential stored at X" → memory classified `context` with `ttl_hours=168`
• Existing context memories (13) are unaffected by the change
• The unreachable TTL branch at harvest.py:753 is now exercised by at least one test

Dependencies

• Blocks: Improvement: add observation→learning synthesis pipeline; strengthen harvest dedup #36 (synthesis pipeline relies on clean type routing — observations should synthesise to learnings; ephemeral facts should route to context, not observation).
• Depends on: none. One-line fix.

Sequencing

Phase 1 (quick wins). Smallest fix on the board — ship first.