[joelteply]

This is plan line 2: a mind's context as an append-only epoch.

Problem. At a turnover, fit_view restarted every fully granted source whole, so the new epoch had no headroom. In a busy room, the next act's appends pushed the view past its grant and it turned over again. Each new epoch starts at a different oldest turn, so the whole prefix changed every act. View::appended_per_act was meant to provide this reserve, and the core never called it.

Change. In prompt/epoch.rs, fit_view and the new start_grants:

• Grounding (above the lowest priority) starts whole.
• The growing sources (conversation, results, thoughts) start at min((A-G)/2, A-G-per_act), where A is the view's grant plus room and G is grounding.
• Allocation. flex_allocate spends the start, collapsing the oldest items first. The newest turn is always kept whole.
• Why half. If the growing sources start at S and keep a reserve R, the epoch lives R/a acts, so re-prefilling costs S·a/R per act. Holding that to at most one act's appends needs R ≥ S, and the measured append rate a cancels out of the bound.
• Before any act is measured, the start is half.

Replay validator. continuum-context/examples/replay.rs was extended to model the window budget and the turnover rule, with the core's measured characters per token. Using the old rule, it reproduces Atlas's live 3% exactly.

Resident / room	Turnovers	Shared prefix, mean / median	Prefill per act, median chars
Atlas 11c1a7ac, before	6	5% / 3%	263,680
Atlas 11c1a7ac, after	1	75% / 89%	26,611
Aris 3be59578 (quiet), before / after	1 / 1	59% / 58%	unchanged
Aris 11c1a7ac, before	6	46% / 26%	194,690
Aris 11c1a7ac, after	2	62% / 77%	53,847

Tests. Two new tests:

• A sliding room used to turn over 30 times in 30 acts; it now turns over at most 2 times.
• At a turnover, grounding stays whole, the oldest turn collapses, and the newest stays whole.

Core: 6,133 passed. continuum-context: 13 passed. Ratchets are unchanged.

Not fixed here (follow-ups, owner Fable):

1. fit_view measures raw tokens while the final fit measures calibrated ones, roughly 3.5% apart. The final fit then cuts the oldest 96–99 messages from the view's head (Benchy). That is a head clip, which goal 7 forbids.
2. Act and deliberation calls share one view with different grants, so switching between them turns the view over.

Live receipt: after deploy, a 15-minute read on the M5 of prefix reuse and act time. Owner Fable; this PR stays open until it is posted.

🤖 Generated with Claude Code

https://claude.ai/code/session_01LoTjvf5j3Ez13g6k8mRkFo

[joelteply]

I removed the unit-mismatch commit (ae895ff) from this PR. It failed provider_context_feedback_bounds_replay_and_scales_next_turn (window_feedback.rs:566), and my first explanation of that failure didn't hold up. It now lives on its own branch, fable/epoch-unit-charge, and will come back as a separate PR once that test passes. This PR is the growth reserve only: head = 5cac179 plus a revert that restores the tree exactly.

[joelteply]

Live receipt, M5, bb91bcf, 05:09–06:40Z. Prefix reuse on warm acts: 88–94%: Aiko 108,972/115,845; Aris 70,392/77,551 and 70,392/80,448. It was 0% before the deploy. Engine prefill on those warm acts was 133–251 s. Acts are still long because of slot waits of 812–2,084 s: every resident went cold at the restart, and each cold prefill of 100k+ tokens held a slot for 1,386–2,611 s. The remaining cost is admission and restart coldness, not the epoch view.