fix(context): protect human anchors and recover current reminders

[Brian Krabach (bkrabach)]

Problem

Compaction treated persisted ephemeral reminders as ordinary user anchors. A reminder could therefore displace the actual first/last human request, and a reminder re-emitted for the current turn could be absent after compaction.

Change

• Select first and last human user anchors using trusted persisted-reminder metadata, not reminder text or the user role alone. The level-8 extreme-compaction fallback retains its established legacy behavior.
• Add an optional, budgeted request-overlay API. When a required trusted reminder is not fully visible, it is restored once at its current pre_user or tail placement for that request only. Canonical history is not duplicated; unsafe placement or an insufficient budget fails visibly.
• The existing no-recovery request path is unchanged. This introduces no blanket pinning, configuration, or lifetime framework.

This context PR independently fixes first-human protection. The overlay API is intentionally present for its companion loop consumer, which remains local and unpublished until this upstream context change is merged; this PR does not claim both changes have shipped.

Verification

• Local unit suites on the existing locks with DTU core 1.6.1: context 146 passed, 1 expected xfailed; loop companion 277 passed.
• Paired baseline-fail/candidate-pass L5 test using the same 32 × 70k synthetic-assistant stress: first human preserved; current reminder exact-once; canonical reminder count one. Also covered current and bundled-system controls, tail expiry, SET_MESSAGES restore/dedup (not full-click resume), and A→B→A supersession.
• Component provider checks (not a full CLI/all-bundle-hooks run): Terra 2/2 returned the marker exactly with policy once on the wire and canonical count one. Sonnet real HTTP accepted policy once with canonical count one and unchanged settings, but 1/1 did not return the exact marker. Only hashes and lengths of returned text were retained, so the cause is unknown; source review confirms this is not a thinking-block parsing issue.
• An earlier Sonnet 400 against a long synthetic-filler fixture is a separate test-limit observation, not a reproduced production 400.

Outstanding gates

• Draft / not merge-ready: investigate the Sonnet behavioral failure before merge. Green CI would not resolve this gate.
• No new full benchmark or weakened assertion is proposed to bypass it.
• The companion loop commit stays local and unpublished pending upstream merge order.