feat: implement durable hybrid OCR execution path

[Sam123ben]

Objective

Deliver the production-grade, reusable OCR execution path behind the existing provider-neutral API: authoritative immutable source binding, digital-PDF text extraction, bounded Rust preprocessing/local inference, confidence-based managed OCR fallback, normalized evidence output, deterministic validation, and durable Temporal orchestration.

Architecture decisions

• Rust/Axum remains the public API and job-control boundary.
• CNPG is authoritative for identity, workflow checkpoints, idempotency, outbox and audit metadata; GCS retains immutable documents/page artifacts/results; Valkey is a tenant-scoped non-authoritative cache/admission layer; Qdrant contains rebuildable derived memory only.
• Before image OCR, the parser extracts embedded text from eligible digital PDFs with source coordinates and provenance. Pages requiring raster OCR follow bounded decode, quality assessment, orientation/geometry correction and local signed-model inference.
• A managed Document AI provider is a policy-controlled fallback for low-confidence, handwriting, unsupported-script or complex-layout pages only. It is never invoked with caller-controlled provider URLs/credentials and is subject to residency, health, cost and tenant policy.
• A dedicated Temporal namespace named document-intelligence-<environment> is mandatory. Workflow histories and inputs remain identifiers-only; source bytes, OCR text, signed URLs and secrets never enter Temporal.
• Model runtime stays provider/model-neutral and Rust-native. No third-party OCR source/runtime is copied or named in repository artifacts. Model artifacts must be legally approved, signed, digest-pinned and benchmark-promoted before use.

Delivery slices

• Persist/load immutable accepted source locator scoped to product, tenant and job; prove cross-tenant/object substitution fails closed.
• Add bounded source-page rendering/extraction adapter and page-source contract; preserve original page geometry.
• Implement concrete signed-model detector/recognizer runtime, bounded CPU/GPU pools and output contract validation after artifact approval.
• Add confidence/calibration routing and managed fallback with circuit breaker, per-tenant quota and cost telemetry.
• Wire checkpointed pages into Temporal activities/finalization; qualify cancellation, retries, 300-page partial recovery, replay/versioning and namespace isolation.
• Add sandbox synthetic/adversarial fixtures, golden-dataset benchmarks, DevAI trace/evaluation gates and redaction assertions.
• Add reviewed GitOps sandbox workload and Kargo promotion gate only after the preceding runtime/evaluation evidence passes.

Acceptance evidence

A disposable sandbox document must complete upload inspection, immutable promotion, parser/text extraction or rasterization, recognition/fallback, evidence/result artifact, validation and redacted trace. Required adversarial coverage: cross-tenant source substitution, stale object generation, decompression/pixel limits, invalid model output, malformed provider response and prompt-injection text. Promotion requires agreed CER/WER, critical-field, latency, memory, cost and safety thresholds.

Explicit non-goals

This issue does not select model weights, create/rotate secrets, alter production Temporal namespaces, or authorize a production rollout. Those actions remain separately reviewed and gated.

[Sam123ben]

Merged PR #48: signed model-profile admission gate. The full CI, remote container contract, OCI image build/scan/smoke, database coverage, and secret scans passed. This advances the prerequisite for safe OCR inference; it does not authorize a workload rollout before an approved signed model artifact and benchmark evidence exist.

[Sam123ben]

Correction to the previous update: PR #48 was a duplicate and has been closed. The signed model-profile admission gate was already merged by PR #40 (0c423eb). The reported CI evidence remains valid; no duplicate code was merged and no workload rollout has been performed.

[Sam123ben]

Tracked the managed-provider infrastructure prerequisite separately in #49. Read-only GCP discovery established that Document AI is currently disabled; no provider or IAM change has been applied.

[Sam123ben]

Merged PR #50: the accepted immutable source now carries a bounded, contiguous page-geometry plan. Parser process, CNPG replay, immutable-source validation and direct accepted-source fixtures all enforce exact page/count/pixel consistency. Local Rust CLI checks and the full remote quality, database coverage, image build/scan, secret and container-contract suite passed.

[Sam123ben]

Post-merge evidence: release image run 33878264866 succeeded and advanced deploy to 4094dd8; post-merge CI run 33878265587 passed Rust quality, full database coverage, both image build/smoke/scan jobs, secret scan and GitGuardian. Read-only GKE check: both document-intelligence Argo applications are Synced/Healthy, but the document-intelligence namespace has no Deployment, Service or Pod and Kargo sandbox/prod stages have no current Freight. There is therefore no API endpoint to invoke until the recognizer/provider and reviewed workload manifests are completed.

[Sam123ben]

Current review evidence and next slice:

• Parent prerequisites from feat(ocr): bind accepted sources durably #47 and feat(ocr): persist bounded page geometry #50 are present on main: accepted GCS source binding is scoped and generation/digest/page-geometry checked.
• The remaining execution gap is now tracked explicitly in feat(ocr): wire accepted sources into a concrete page recognizer #51: PageRecognizer is an interface with only test implementations, ocr-service does not construct a page-recognition worker, and parser inspection returns page geometry rather than OCR observations.
• Read-only production verification on 2026-09-04: project tesseracthub-480811, context gke_tesseracthub-480811_asia-south1_tesseract-prod-in-gke, namespace document-intelligence. Both Argo applications are Synced/Healthy; the namespace has no Deployment, Service or Pod; sandbox/prod Kargo stages have no current Freight.

This keeps the issue status honest: there is no usable OCR endpoint to test or deploy until #51 has a concrete signed local recognizer or #49 provisions the managed fallback, followed by the reviewed GitOps workload and promotion evidence.