Verify OMP preset activation and cross-provider coexistence after native-history fixes #205

Problem

Native OMP v3 ingestion and its preset are delivered by #124 / PR #125. Native preset installation is delivered by #123 / PR #188. Read-only inspection found an active OMP manifest in the local runtime configuration; actual installed-revision ingestion and recall remain unverified. One integrated proof must establish that OMP works alongside the corrections in #199, #200, #201 and #159.

Mixed-fixture verification on clean integrated revision d77d62e exposed a recovery defect: planning rejects canonical identities emitted by the native OMP and Pi readers because compat.ValidateRecoveryUUID accepts only RFC UUIDs and Claude-derived suffixes. The actual CLI exited 1 before recovery mutation. The correction must honor existing canonical opaque reader identities and preserve collision, payload and ownership checks; it adds no rich-event recovery guarantee. Keep a consumer-visible regression and rerun the same mixed recovery path after the fix.

Change

Exercise the shipped OMP reader and preset on the integrated revision after the open prerequisites merge. First ingest a synthetic mixed-provider fixture through the actual CLI in isolated HOME/config/SQLite state, including OMP, Pi, Codex and two native OpenCode sessions/projects in one source database. Then consume a genuine OMP headless source using the clean candidate development binary installed into a private prefix with fresh private HOME/configuration/index state. Preserve an already active manifest; install missing presets through presets install --config-home with the explicit runtime configuration home. Obtain exact-target approval before modifying an existing manifest or using the six-preset overwrite operation. Public evidence contains synthetic data or sanitized counts only.

Scope is supported activation, a consumer-visible mixed-provider integration regression, and observed coexistence. Existing source, privacy, origin, replay, canonical ownership and derived-metadata recovery contracts remain authoritative.

Acceptance

1. The actual CLI installs/preserves the shipped presets and ingests the deterministic mixed fixture. Record revision, effective isolated configuration and expected provider counts. OMP discovery uses the canonical direct-session JSONL shape and native v3 decoder.
2. OMP preserves header/cwd identity, native physical-block UUIDs and parser-backed origins. Reasoning remains opt-in; toggling it retains IDs, ordinals and tool coordinates. Appended results enrich native error/exit/interruption signals without duplicate calls. Commands, failures, sequences and existing filtered/trend consumers retain proven and unknown outcomes correctly. Direct-search echo eligibility remains correct.
3. Pi and Codex each emit one event per supported native call, pair interleaved outputs by native ID, preserve explicit success/failure/unknown and replay unchanged available sources through the shared bounded scheduler. Codex wrapper success remains distinct from nested-operation outcome. OMP native ToolUseID populates nullable tool_events.native_call_id while message_uuid remains the canonical message owner. Existing recovery rederives tool_events: available inputs can restore native metadata; expired-source lossy rederivation retains unknown native IDs. Verify that boundary without adding a new rich-event recovery contract.
4. More than 200 eligible matches per FTS stream demonstrate the adopted bounded search contract, stable ranking, scope and echo filtering. Candidate exhaustion is truthful on early and empty later pages; typed search retains direct pagination. JSON retains its array shape. Robot candidate/omission metadata respects the complete escaped token budget, including tiny-budget stderr fallback.
5. Two OpenCode sessions/projects in one source remain separate in list, search, exact context, tags and sequences. Context stays in the owning session; sequences and project/tag scopes remain isolated. Updates, replay and canonical recovery preserve native ownership and retained payload.
6. After isolated mixed-fixture PASS, prove consumption by the clean candidate development binary installed into a private prefix with fresh private HOME/configuration/index state of a genuine installed OMP 18.12.6 headless session created through headless provider inference with two native bash calls. Record the exact candidate revision, effective private manifest/configuration and target scoping, using canonical direct-session logging. Verify opaque IDs, origin, cwd, recall JSON and robot/context/pattern behavior; pair native bash calls by their native IDs and canonical message owners. Preserve source and production manifest unchanged, and retain the session. Establish isolated mixed-provider PASS separately. Preserve successful native calls whose output omits exit status as success with exit unknown; preserve failure with is_error=true and exit 7. Do not claim reasoning-content indexing without thinking blocks. Observe genuine-provider proof independently of synthetic-fixture success. The original global installed mixed-target acceptance involving OMP/Pi/Codex/Claude/OpenCode is tracked in a separate follow-up created by the same backend run.
7. Run the targeted commands below, the final corrective regressions from the four delivered issues, and the normal full gate once after integration. Confirm the actual merged test names and record exact commands, exit status and observed outcomes. Update existing operator documentation for the activation path actually exercised.

Targeted verification

go test ./internal/readers ./internal/input_config ./internal/storage ./cmd/backscroll -run 'OMP|Pi|Codex|OpenCode|Native|IngestionFingerprint|ReplayScheduler|Search|Robot|Budget|RefillCandidate' -count=1
go test ./internal/presets ./cmd/backscroll -run 'InstallUsesExplicitConfigHome|InstallPreservesExistingPreset|InstallOverwritesWithExplicitOption|EmbeddedPresetsMatchShippedInputs|PresetsInstallUsesExplicitConfigHome' -count=1
just check && just ci

CLI proof must include config, list, search in JSON and robot modes, context using an actually returned UUID, commands/failures/sequences, repeated startup, appended results and reasoning changes. Use fixture-specific expected counts and ownership, not exit zero alone. The cap fixture includes both first and beyond-pool pages plus typed contrast. All initial proof uses private temporary HOME/config/database state and a dev binary from the integrated revision. Local operational commands can synchronize the index and require target authorization.

Dependencies and delivery provenance

All four corrective prerequisites are merged and closed: #199 / PR #204 (Pi native events), #200 / PR #207 (Codex native events/V19), #201 / PR #210 (truthful bounded pagination), and #159 / PR #209 (OpenCode native session/project ownership/V20). Their native feature runs succeeded and completed owned-resource cleanup. The three native blocked-by relationships identify these completed prerequisites; they are no longer unresolved blockers.

Integrated main is d77d62e. Five separate isolated real-CLI E2E cases passed against a clean dev binary from that revision. They establish per-provider delivery evidence, not the single mixed-index proof or authorized installed-target proof required by this issue. #206 / PR #208 delivered the omp-dagu 0.3.94 adoption; its native run also succeeded and cleaned its resources. Runtime investigation remains separate from this integration scope.

Existing OMP v3 ingestion (#124 / PR #125) and native preset installation (#123 / PR #188) are completed provenance. This issue verifies their activation and coexistence rather than implementing a second reader. It remains open until the acceptance criteria above, including genuine OMP headless/private candidate-consumer proof, are observed.

Breakdown source: plan

Operator-authorized scope split: close this bounded isolated-mixed plus actual OMP-headless consumer proof; preserve the original global installed mixed-target acceptance in the separate follow-up. Production inputs and manifests remain unchanged. Public headless evidence: #205 (comment)