DeepSeek session at 842k/1M never auto-compacted — investigate trigger gap

[Hmbown]

User report 2026-08-23: a DeepSeek session shows 842k/1M context and auto-compaction never fired (threshold default 80% = 800k on a 1M window).

Investigation prompt with ranked hypotheses + instrument plan: codewhale-ops/TAKEOVER-PROMPT-COMPACTION-20260823.md. T1 (meter/trigger alignment) already landed on the 0.9.12 branch; the report may be the pre-T1 inflated display, or the real trigger gap (billed-token wiring / bounded-fragment mismatch / reclaimability floor).

Acceptance: root-cause with exact numbers, fix the real gap, regression test pinning 'a 842k-billed 1M-window session must compact'.

[Hmbown]

Fixed on codex/v0912-integration-20260823 in fda5d245f — regression billed_842k_on_a_1m_window_compacts_despite_a_small_estimate pins the 842k/1M compact trigger.