feat(finance): add continuous watch review loop

[ashione]

Summary

• add productized finance intelligence stages for persistent evidence indexes, retriever runtime, durable job control, and replay evaluation
• split finance evidence indexing/search logic into finance_pack.evidence_index and add FinanceEvidenceRetriever so CLI orchestration no longer owns retrieval internals
• remove hash embedding from the finance intelligence product path; semantic retrieval is explicitly disabled until a real production embedding provider is configured, and v2 indexes force stale hash-vector indexes to rebuild
• add finance_intelligence_jobs plus finance intelligence jobs/status/stop/resume so agents can inspect, stop, and resume durable watch/intelligence runs without scraping logs
• add finance_intelligence_evaluations plus finance intelligence evaluate/eval-report for persisted-row signal/provider/retrieval/runtime quality metrics
• update CLI help, Python README, plan/spec docs, and velaria_python_local skill examples so velaria_cli_run can discover the complete workflow

Validation

• TDD red checks observed for missing finance_pack.jobs and missing finance_pack.evaluation, then fixed with focused tests
• uv run --project python python -m unittest python.tests.test_finance_pack python.tests.test_python_cli_contract -> 100 tests passed
• bazel test --cache_test_results=no //python:finance_pack_test //python:python_cli_contract_test //:python_ecosystem_regression -> 14 targets passed
• uv run --project python python python/velaria_cli.py finance intelligence --help -> includes jobs, status, stop, resume, evaluate, and eval-report
• git diff --check -> passed
• focused changed-file secret scan -> no real secrets found
• live persisted-session status smoke with VELARIA_HOME=/tmp/velaria-finance-reviewfix-us-20260520: finance intelligence status --session-id us_watch_reviewfix_20260520 --format json returned event_count=61, signal_count=4, and effective_status=completed
• live retrieval smoke on the same session: finance intelligence search --query "AAPL MSFT NVDA realtime news fundamentals signal" --format json returned retriever=finance_evidence_retriever, index_status=hit, semantic.status=disabled, doc_count=61, and top symbol AAPL
• live durable job smoke on the same session: finance intelligence jobs --session-id us_watch_reviewfix_20260520 --format json returned the persisted evidence_search_job under finance_intelligence_jobs
• live replay evaluation smoke on the same session: finance intelligence evaluate --session-id us_watch_reviewfix_20260520 --format json returned quality_status=ok, row_count=61, signal_count=4, search_count=9, job_count=1, and provider_refetch=false
• live evaluation report smoke on the same session: finance intelligence eval-report --session-id us_watch_reviewfix_20260520 --format json returned quality_status=ok, no_hash_embedding=true, and research-only disclaimers
• GitHub PR checks for latest commit 9f2856a: native-and-python passed, python-wrapper-leak-smoke passed, wheel jobs skipped by workflow

Notes

• Outputs remain research evidence only, not investment advice.
• Replay/evaluation commands read persisted rows and durable job/search/index events only; they do not refetch public providers.
• Quote freshness is provider-reported; Yahoo U.S. quotes may be delayed.
• Semantic retrieval remains disabled until a production embedding provider is explicitly configured.

Review follow-up: a318d4a

• Hardened finance intelligence resume so it only executes trusted persisted watch-session start argv records with Python/module/session markers and matching argv fingerprint; untrusted records now return structured resume_unavailable and do not call Popen.
• Fixed text-mode finance intelligence resume when the runtime is already running so it renders the normal intelligence report instead of exiting silently.
• Extended replay evaluation hash checks to detect legacy vector/hash artifacts such as vectors_path, vector_path, vectors_json, vector_count, and nested semantic provider metadata.
• Added regression tests for all three review findings.

Review follow-up validation:

• Focused red/green tests for the three findings: 4 tests passed after fixes.
• uv run --project python python -m unittest python.tests.test_finance_pack python.tests.test_python_cli_contract -> 103 tests passed.
• bazel test --cache_test_results=no //python:finance_pack_test //python:python_cli_contract_test //:python_ecosystem_regression -> 14 targets passed.
• git diff --check -> passed.
• focused changed-file secret scan -> no secrets found.
• live persisted-session evaluation smoke now detects legacy hash-vector artifacts in the older /tmp/velaria-finance-reviewfix-us-20260520 session: quality_status=failed, no_hash_embedding=false, provider_refetch=false.

Review follow-up: 730c620

• Added a structured SEC Company Facts provider policy via sec_user_agent_policy(), replacing the previous placeholder contact User-Agent with an environment-configurable VELARIA_SEC_USER_AGENT policy.
• Wired finance doctor to report sec_companyfacts_policy in JSON/text flows so agents and users can see whether SEC access is production-configured.
• Updated Python README and velaria_python_local skill examples to tell users to set a descriptive SEC User-Agent before production fundamentals fetches.

Review follow-up validation:

• TDD red check observed before implementation: focused tests failed because sec_user_agent_policy was missing.
• Focused red/green tests for SEC policy/fetch/doctor: 3 tests passed.
• uv run --project python python -m unittest python.tests.test_finance_pack python.tests.test_python_cli_contract -> 105 tests passed.
• bazel test --cache_test_results=no //python:finance_pack_test //python:python_cli_contract_test //:python_ecosystem_regression -> 14 targets passed.
• uv run --project python python python/velaria_cli.py finance doctor --skip-network --format json -> reports sec_companyfacts_policy warning with default non-production policy.
• VELARIA_SEC_USER_AGENT=... finance doctor --skip-network --format json -> reports sec_companyfacts_policy ok with configured env source.
• git diff --check -> passed.
• focused diff secret scan -> no secret-pattern hits in added lines.

Information-source follow-up: 75def60

• Added two more public, no-key information sources: yahoo-finance-news for symbol-focused Yahoo Finance RSS headlines and sec-filings for official SEC EDGAR Atom filing events.
• Added provider/row-level source classification and scoring: source_category, source_type, source_score, and source_score_reason now appear in finance sources catalog output and fetched evidence rows.
• Fed source_score into finance evidence structured ranking so higher-authority sources can influence hybrid search ordering alongside keyword, symbol, signal, and recency signals.
• Updated Python README and velaria_python_local skill examples so agents can discover and run the new sources directly.

Information-source validation:

• TDD red check observed before implementation: focused tests failed because classify_information_source was missing.
• Focused red/green tests for provider catalog, source classification, Yahoo RSS parsing, SEC Atom parsing, and source-score ranking: 5 tests passed.
• uv run --project python python -m unittest python.tests.test_finance_pack python.tests.test_python_cli_contract -> 109 tests passed.
• bazel test --cache_test_results=no //python:finance_pack_test //python:python_cli_contract_test //:python_ecosystem_regression -> 14 targets passed.
• uv run --project python python python/velaria_cli.py finance sources --format json -> catalog includes source category/type/score fields.
• Live finance fetch-news --provider yahoo-finance-news --market us --symbol AAPL --limit 2 -> returned 2 scored news rows.
• Live VELARIA_SEC_USER_AGENT=... finance fetch-news --provider sec-filings --market us --symbol AAPL --limit 2 -> returned SEC 10-Q/8-K filing rows with source_score=0.96.
• git diff --check -> passed.
• focused diff secret scan -> no secret-pattern hits in added lines.

Code-review follow-up: f9e47c9

• Fixed Yahoo chart/history rows so they are enriched with source_category, source_type, source_score, and source_score_reason like other fetched evidence rows.
• Wrapped sec-filings SEC ticker-map failures in FinanceProviderError so CLI/agent callers receive structured provider diagnostics instead of raw exceptions.
• Made SEC User-Agent policy identify the shared SEC policy as provider=sec because it is used by both sec-companyfacts and sec-filings.
• Clarified fetch-news --query help in both the direct finance pack parser and the public velaria finance wrapper.

Code-review follow-up validation:

• Focused red tests reproduced the review gaps before implementation.
• Focused review-fix tests plus doctor smoke: 4 tests passed.
• Public CLI help smoke: finance fetch-news --help now documents provider-specific query semantics.
• uv run --project python python -m unittest python.tests.test_finance_pack python.tests.test_python_cli_contract -> 110 tests passed.
• bazel test --cache_test_results=no //python:finance_pack_test //python:python_cli_contract_test //:python_ecosystem_regression -> 14 targets passed.
• git diff --check -> passed.
• focused diff secret scan -> no secret-pattern hits in added lines.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 2c0d270975

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Avoid rescanning all events in each supervise iteration

_build_watch_session_review computes summary via _summarize_watch_session(session), which already performs a full _read_watch_session_events(..., limit=0) scan, and then immediately performs another full scan for rows on the next line. In watch-session supervise (especially long-running --iterations 0 sessions), this doubles per-iteration I/O and makes review latency grow with session size, so the supervisor can fall behind and produce increasingly stale diagnostics.

Useful? React with 👍 / 👎.