feat: multi-provider foundation (SessionRef extension + provider pricers)

[0bserver07]

Summary

Implements Steps 1 + 2 of docs/specs/multi-provider/spec.md — the prerequisite work so Wave 2 (Cursor + Cline adapters) can land without further refactoring of the adapter contract or cost layer. No new providers in this PR.

§1 — Adapter contract extension

• SessionRef gains optional source_kind: Literal["file", "database"] = "file" and source_hint: dict[str, Any] | None = None. JSONL adapters need zero changes.
• ingest_log migrated from file_path PRIMARY KEY to (id INTEGER PRIMARY KEY, file_path, session_id TEXT, storage_kind TEXT CHECK(...), mtime, size, processed_offset NULLABLE, last_rowid NULLABLE, last_ingest_ts) plus two partial unique indexes — one for file-mode rows (session_id IS NULL), one for database-mode rows. Existing rows survive with session_id=NULL, storage_kind='file', last_rowid=NULL. Migration is via CREATE … INSERT SELECT … DROP … RENAME inside one transaction; idempotent because apply() reads PRAGMA user_version and skips already-applied migrations.
• run_ingest() branches on ref.source_kind: file-mode resumes via processed_offset keyed on WHERE session_id IS NULL; database-mode resumes via last_rowid keyed on (file_path, session_id).
• Writer stores max(record.seq) for both kinds — for database mode that's a rowid, for file mode it's the byte offset of the last consumed line. Adapters skip records with seq <= since_offset so the storage-aware contract test (test_read_since_offset_is_storage_aware, spec §1.4) passes for both providers.

§2 — Provider pricer scaffold

• New stackunderflow/infra/providers/ package: ProviderPricer ABC plus AnthropicPricer and OpenAIPricer extracted from the old infra/costs.py heuristic. get_pricer(provider) returns a singleton, falls back to Anthropic for unknown providers.
• infra/costs.compute_cost(tokens, model, provider="anthropic") is now a thin shim — the default keeps every existing call site working unchanged. Aggregator collectors take a provider= constructor arg resolved once from ds.records[0].provider; records carry it through from RawEntry/TaggedEntry populated from projects.provider in build_enriched_dataset.
• Codex normalization moved out of adapters/codex.py into OpenAIPricer.normalize_tokens. The adapter now delegates to the pricer instead of inlining the cached-input subtraction + reasoning fold. A parametrised regression test (tests/stackunderflow/infra/providers/test_codex_cost_equivalence.py) proves the move is cost-neutral by computing the same (model, raw-token) bundle with both the legacy convention (caller pre-normalises) and the new one (pricer normalises) — they match for 5 fixtures including a 100%-cached-input edge case.

Conservative choice (not in spec)

The spec wanted adapters/codex.py to emit raw OpenAI shape into Record. The 4-slot Record dataclass (input_tokens, output_tokens, cache_create_tokens, cache_read_tokens) doesn't have a clean place for reasoning_output_tokens to live without re-purposing cache_create_tokens (which would corrupt downstream cache stats). I chose the smallest-diff path: the adapter still emits 4-slot canonical shape into the DB, but the flattening logic now physically lives in OpenAIPricer.normalize_tokens — the adapter calls into the pricer instead of inlining the math. The cost-equivalence regression test guarantees the dollar number doesn't change. A future PR can switch the adapter to raw shape once Record grows a fifth slot or the encoding question gets resolved.

Test plan

• pytest tests/ -q — 484 passed, 2 skipped (baseline was 444 + 2; +40 covering the new pricers, the contract test, the v002 migration, and the cost-equivalence regression).
• pytest tests/stackunderflow/adapters/contract.py tests/stackunderflow/adapters/test_claude.py tests/stackunderflow/adapters/test_codex.py -q — 29 passed.
• cd stackunderflow-ui && npm run typecheck — clean.
• cd stackunderflow-ui && npm run build — clean.
• Smoke-ingest against a local Codex project to verify the v002 migration applies cleanly to a real ~/.stackunderflow/store.db.

🤖 Generated with Claude Code