Combine exact outcome evidence across participating machines #106

What: Combine exact outcome evidence from participating machines without billing whole sessions.
Why: Existing fleet task figures pool session usage and its sanitizer drops unversioned extra fields.
So what: After reviewed local accounting exists, one exclusive implementation versions caller-named exact export and union at the existing fleet owner.

Outcome

A caller can export one outcome per machine and obtain a deduplicated exact report with missing machines, usage and rates visible. Existing whole-session and judge evidence remains usable. This advances Observer's cross-machine native-evidence Objective without moving transcripts.

Acceptance criteria

• New schema is a superset of existing session/task/judge fields plus one caller-named exact section; version producer, validator, sanitizer, storable/render checks, store/read, CLI selection and exact union together. Unsupported schema2 cannot become exact spend.
• Use existing single machine slot/local precedence and caller-invoked transport. Display cutoffs and missing/different-task/stale scope; no archive, all-task exact scan or new transport.
• Union original harness/native response identities once across copied ledgers and changed local aliases; contradictory purpose/final counters/evidence conflict rather than first/latest wins. Preserve native counter semantics and price one union under one sourced schedule.
• Reconcile expected remote launch/child membership against compatible exports. Unknown source/index/participation or native finality is a gap, never zero. SSH-control-only model usage stays on its actual agent origin.
• Updated Observer skill and applicable AgentsMD guidance use the approved versioned purpose and exact report contracts. Keep shared terminology at its owner and do not promise historical backfill or full costs without proof.

Blocked by

#105, until its complete independently reviewed candidate and exportable evidence contract are recorded. Native source377 is a transitive dependency.

Required proof

Round-trip real exact fields through validate/sanitize/store/read/union, preserving whole-session/judge evidence; A and C in one session; copied response dedupe; incompatible/old schema; missing remote; older/different-task slot; pull failure; conflicting counters/purpose; unknown prices and origin. Separate native verification and actual participating-machine real-task proof with honest gaps. Focused tests, current CI, independent exact-head review and sanitized report inspection. No unsupported generated report is published as accepted full cost.

Parent

#58

Non-goals

No new transport, daemon, rates, manual ownership, invented historical purpose, subscription allocation, purpose repair or production deployment.

Elon record

Requirements and who asked: User approved the multiple-Issue accounting rule and requested building it; parent58 records the reviewed policy and native proof requirements.
Deleted: Whole-session substitutes, opaque registry, duplicate evaluator, manual capture and false full costs.
Bottleneck: Reviewed native membership/finality evidence at existing owners before downstream aggregation.
Checked myself: Live Observer main71bcb597a3ec8a15f3c595d043fc07d0e2271b59, native Chromeria3caa source, current report/fleet owners and adjacent72; required runtime proof is not yet passed.