[long-run] Add versioned retained-resource and semantic-yield interval telemetry

[chaoz23]

Parent: #215. Extends the shipped #98 accounting; does not reopen #98.

Product outcome

Let a creator or agent see what a long InkCheck run retains, what useful evidence each interval adds, and whether post-discovery work is converting memory and time into author-facing value. Keep raw throughput separate from semantic yield.

Contract

• Add a versioned bounded-interval sample with deterministic transition position and observational elapsed/process-resource fields.
• Attribute current and peak counts/bytes to pending and active state payload, exact dedupe, ancestry/witnesses, semantic indexes, frontier references/node slots, findings, checkpoint buffers, and finalization reserve.
• Define component ownership so shared backing is charged once; report deterministic logical estimates separately from V8 heap/RSS/external memory and an observational unattributed remainder.
• Record a yield vector, not one score: critical findings, approved intent/stages, authored locations/choice edges, visible-outcome fallback, bounded meaningful variable transitions, exact terminal variants, and raw state territory.
• Distinguish first useful/critical discovery from post-first-discovery yield, including bounded dry-gap summaries and campaign-new versus rediscovered identities.
• Keep progress and compact MCP output privacy-safe and bounded; full detail remains drill-down data.
• A policy decision is a pure function of persisted sample inputs. Wall time may bind a ceiling but must not silently perturb frontier order.

Acceptance criteria

• Samples appear at a fixed transition cadence, epoch/checkpoint boundaries, pressure actions, and termination.
• Every major retained owner has current/peak count and byte fields; totals are internally consistent and clearly labeled as logical estimates or observed process metrics.
• Visible outcomes, assertion violations, goals, and stages can independently trigger the interval record that contains them.
• Reports distinguish early discovery from post-discovery yield and expose category-specific per-million-transition and retained-GiB-minute inputs without an opaque aggregate score.
• Identical source/config/seed/policy/sample ledger produces identical structural actions and semantic identities.
• Tests cover zero-yield, late recovery, assertion/goal boundaries, large finding sets, logical-versus-RSS separation, sample compaction, and privacy/response bounds.
• Calibration records telemetry overhead; promotion requires no more than 5% wall/RSS overhead or explicit accounting.

Non-goals

• Activating shadow long-tail allocation.
• Treating terminal multiplicity or raw state novelty as actionable value.
• Choosing checkpoint encoding or next-window size.

[chaoz23]

Acceptance evidence map — 2026-08-25

• PR Add bounded shared-search observability #224 added fixed-cadence/termination samples and the initial bounded semantic-yield ledger.
• PR Add deterministic shared observability boundaries #225 added deterministic category/frontier boundaries, timing, migration, bounded live output, and integrity-bound observability-v2 checkpoint state.
• PR Account for checkpoint and finalization buffer ownership #226 added deterministic retained-search plus non-recursive ledger current/simultaneous-peak ownership, unallocated finalization reserve, and checkpoint/report write receipts.
• PR Account for checkpoint reopen and report enrichment ownership #227 added checkpoint reopen/resume and report-enrichment/finalization lifecycle receipts. Exact local suite: 351 total / 348 pass / 0 fail / 3 expected sandbox skips; pre-merge 10/10 and post-main run 32896092156 5/5.

Still missing: emitted checkpoint/epoch/pressure boundaries, complete shared-backing-once owners (including search-session lifecycle), category per-million and retained-GiB-minute inputs, campaign-new versus rediscovered summaries, remaining kth/throughput facts, and the final combined overhead/policy-neutrality gate. The current median calibration is narrow engineering evidence, not a promotion result. This issue remains partial and open.