fix: context health note — signal omitted details instead of false assurance (#434)

## Summary

Updates the context health note injected at gradient layer 1+ to
honestly signal that distilled observations are lossy summaries with
details omitted, rather than falsely assuring the model that details are
preserved.

## Problem

The context health note said:

> Earlier details (file paths, error messages, decisions) are preserved
in distilled observations and searchable via the recall tool.

This gave the model false confidence. When it saw a compressed
distillation mentioning "flock locking" but not the rejected "staleness
check" alternative, it answered "no alternatives were considered"
without invoking recall — because the note said details were preserved.

## Fix

Updated both variants of the context health note to:

> Distilled observations above are lossy summaries — specific details
(exact error messages, rejected alternatives, file paths, numerical
values) are likely omitted. Use recall to verify any specific claim
before answering questions about what happened, what was considered, or
what the exact values were.

Key changes:
- "preserved" → "lossy summaries" — honest about compression loss
- Lists specific detail types that are likely omitted (rejected
alternatives, exact error messages)
- "Use recall proactively" → "Use recall to verify any specific claim" —
stronger guidance

## Eval Results (CM-1, 400K inflation)

Two runs: **4.62** and **4.69** (average ~4.65, up from 3.69 baseline).

m3 (the persistent failure) scored **5.0** on one run when the
distillation observer preserved the alternative, and 2.2 on the other
when it didn't. The remaining variance is in distillation quality
(non-deterministic LLM output), not in the context health note.

## Tests
- 1752 pass, 0 fail
- Typecheck clean across all 4 packages