agents-efficiency-KPI v2: settle the actuator coverage gap and make the report recurring

v1 shipped 2026-09-11 at /data-venus/projects/savetoke/reports/2026-09-11-agent-efficiency-v1.md and was corrected three times; what it could NOT answer is carried here so it is not re-derived. 1. ACTUATOR COVERAGE (the one real gap). v1's claim that the actuator fires and clears contexts is confirmed against event_log for the mars 400k seats ONLY - coder-mars-cc, coder02-mars-cc, db-mars-cc at 294k/320k, 8-9 compactions each over 7d, verified by monitor-context-cc. Every 250k seat topping finding 2's count table is UNCHECKED, so for them 'the actuator cleared this context' is an inference. Fix: one event_log query on monitor-context-cc's surface covering db-terra-cc (9 obs above its compact tier), bin-venus-cc (7), coder-venus-cc (5), coder-pluto-cc (4), audit-terra-cc (4), db-pluto-cc (4), coder-enamel-cc (4). Owner of the surface: monitor-context-cc. 2. NEVER RE-DERIVE A SHARE FROM context.samples. Three headline numbers were published and withdrawn from this data: 4.0% raw (contaminated by frozen re-emissions), 20.0% dedup (= 1 observation out of 5, a quantization artifact), and the 55-70% median band (its upper edge was coder02-mars-cc's frozen 281,463). Finding 2 now reports integer COUNTS and refuses to convert. v2 must not reopen this - the instrument that answers 'how often does an agent hit its tier' is event_log, which records events and has no denominator to contaminate. 3. STILL MISSING FROM THE INPUT DATA: bin-venus-cc's model/effort capture (v1 shipped without it under the explicit Elazar GO to deliver from what exists today); whey renders as NO COVERAGE and must keep doing so - never as a zero or a good score - until its collector is repointed off the dead token-usage.sqlite default. 4. RECURRING DELIVERY is the epic's actual goal and is not started: v1 is a one-off markdown file. Needs a cadence, a target (Elazar reads the hub, not a file path), and the kpi view fix from sh.git 91eb3e6 to land first or episode counts stay inflated. Headline metric stays turns and standing-context size, not message payload (Elazar, relayed by pmmaster-evolutiva-cc).