Add page-memory VLM capacity gate

[suguanYang]

Summary

• Add a Redis-backed global page-memory VLM in-flight limiter so page-memory capacity pressure becomes retryable UnavailableException backpressure instead of degraded empty OCR/title/summary output.
• Gate only usage_task values prefixed with page_memory. in the shared OpenAI-compatible sync client; non-page-memory LLM calls keep existing behavior.
• Add env-backed page-memory local concurrency defaults for scope processing, page tagging, title detection, and node assembly. Title detection and node OCR/summary now default to concurrency 3; existing scope/tag defaults remain 5/4 but are tunable through settings.
• Add ordered local parallelism for page title detection and node OCR/summary assembly, preserving page/row order and failing stages on greenlet exceptions.
• Add page-memory sub-stage timers for title render, node OCR, node summary, and node rows.
• No public API, SDK, OpenAPI, or DB schema changes. Existing job metadata carries the resolved page_memory_config.

Verification

• uv run pytest packages/shared-python/shared/tests/test_page_memory_config.py packages/shared-python/shared/tests/test_page_memory_vlm_limiter.py apps/worker/tests/contract/test_page_memory_page_tagger_contract.py apps/worker/tests/contract/test_page_memory_node_assembler_contract.py
• uv run pytest apps/worker/tests/contract/test_page_memory_*.py packages/shared-python/shared/tests/test_page_memory_config.py packages/shared-python/shared/tests/test_page_memory_vlm_limiter.py
• uv run pytest apps/worker/tests/contract/test_page_memory_*.py
• uv run pytest apps/worker/tests/contract
• make check
• Local API health: curl -sS http://localhost:5005/health
• Real v2 parse, text-heavy PDF: job_c5bd6e3f7b34, TSLA 35-page PDF, completed done, parse_track=page_memory, result ZIP /tmp/tsla-page-memory-real-parse.zip.
• Real v2 parse, scanned/image-only PDF: job_af2bc76242f3, synthetic 8-page scanned PDF, completed done, produced 8 non-empty page chunks, result ZIP /tmp/scanned-page-memory-real-parse-with-env.zip.
• Worker logs for job_af2bc76242f3 showed page_memory.vlm_gate.acquired for page_memory.title_detection, page_memory.hierarchy, page_memory.tag, and page_memory.node_ocr; observed in-flight counts up to 4. This real parse was run before the final concurrency default bump, so it observed OCR concurrency 2.
• Redis page_memory:vlm:inflight was empty after completion.

Deployment Notes

• New internal settings with defaults: PAGE_MEMORY_VLM_MAX_INFLIGHT=16, PAGE_MEMORY_VLM_LEASE_TTL_SECONDS=600, PAGE_MEMORY_VLM_WAIT_TIMEOUT_SECONDS=120, PAGE_MEMORY_SCOPE_CONCURRENCY=5, PAGE_MEMORY_TAG_CONCURRENCY=4, PAGE_MEMORY_TITLE_DETECTION_CONCURRENCY=3, PAGE_MEMORY_NODE_ASSEMBLY_CONCURRENCY=3.
• Worker processes must have IMAGE_MODEL exported for page-memory page-level VLM tagging/OCR. During local testing, a worker started without apps/worker/.env skipped page VLM calls; restarting with .env sourced exercised the limiter.
• No migrations, queue changes, storage changes, or public contract changes.
• Rollback is the branch revert; stale Redis in-flight counters are bounded by the lease TTL.

Checklist

• Tests were added or updated when behavior changed
• Public docs, examples, or OpenAPI contracts were updated when needed
• Database migrations are idempotent and safe to deploy
• Logs, errors, and validation paths avoid leaking secrets or user data
• The pull request description explains any breaking or user-visible change