emrg: invalidate usage anchor on mid-session model/provider switch

[argszero]

Summary

Fixes#1000 — the usage anchor is not invalidated on a mid-session model/provider switch.

Root cause

The auto-compact usage anchor (emrg/server/daemon.py) is a per-session (real_prompt_tokens, local_estimate) pair, refreshed after every round from the provider's real prompt_tokens. The auto-compact gate projects:

projected = real_anchor + (estimate_now − estimate_anchor)

so the systematic local-estimator undercount (observed 148K est vs 222K real, #946) never accumulates.

The anchor is deliberately invalidated on auto-compact and manual compact — but not in _handle_set_model. After a mid-session /model switch, the anchor still holds the old model's real prompt_tokens, so the projection carries a stale tokenizer baseline into the new model's rounds — the first post-switch round can underestimate badly and the gate may not fire when it should (the exact #946 failure mode the anchor was built to prevent).

Fix

In _handle_set_model, when the API model actually changes:

• drop all usage anchors (the model is global daemon state — every session's anchor is stale),
• mark the drops deliberate (mirror of the post-compact drop) so the fail-LOUD "anchor missing" warning does not false-positive,
• the next round re-anchors from the new provider's real prompt_tokens.

A no-op switch (same API model re-selected) leaves anchors untouched.

Verification

• 2 new unit tests (test_model_switch_invalidates_usage_anchor, test_model_switch_same_model_keeps_anchor); all 12 usage-anchor/model-switch tests pass
• Full suite: 1103 passed, 1 skipped; import check + emrg --help OK
• Agent.md Python test count synced 1102 → 1104 (doc-count guard)
• No daemon-protocol or client changes

[argszero]

Closing as superseded by #1003 (same fix for issue #1000, submitted ~5 min later by the issue author).

Why #1003 wins the consolidation:

• It covers everything this PR does (drop all anchors + mark deliberate drops on real API-model change, no-op switches untouched)
• Plus it clears pending drift windows (_missing_anchor_est) — my version would let a pre-switch anchor-loss window measure old-model-est vs new-model-real and pollute the countable anchor_drift metric
• Cleaner separate marker set (_usage_anchor_dropped_by_switch) + helper method + 4 tests (vs 2 here)

Keeping all three duplicates open would split LGTM votes across PRs and force messy Agent.md doc-count conflict resolution on any double merge. #1003 is fully verified (CI double-green, full suite 1105 passed + 1 skipped, import OK) and has my LGTM. Merging #1003 closes #1000.

[argszero]

Superseded by #1003 — see comment above. Fix tracked via #1003.