test(checkpoint): measure Postgres storage growth - #5051

Refs #4189

Why

The retention design in #4189 is intentionally blocked on a protected-set/deletion contract. Current coverage measures memory and SQLite growth, while the production Postgres saver has no equivalent full/delta storage-growth row. That leaves the retention discussion without comparable checkpoint/blob/write evidence across durable backends.

This PR is measurement-only. It does not define a retention policy, delete production history, or claim that any measured row is safe to prune.

What changed

• bench_channels.py accepts an opt-in postgres backend when TEST_POSTGRES_URI is set and runs it through AsyncPostgresSaver.
• Every Postgres case uses a unique benchmark thread, verifies warm and reopened materialization, then removes only that temporary thread through the saver's public adelete_thread API.
• Versioned JSONL schema v2 reports normalized checkpoint_*, blob_*, and write_* row/byte fields for memory, SQLite, and Postgres, while retaining aggregate logical byte fields.
• SQLite reports zero separate blob rows because SqliteSaver embeds channel values in its checkpoint payload; Postgres reports checkpoint_blobs separately. The docs call out that backend byte encodings are not directly identical.
• The summarizer includes the new storage fields by default.
• Unit coverage pins the normalized Postgres queries, URI opt-in, case-scoped shared-backend thread IDs, diagnostic degradation, and memory/SQLite storage shapes. A real Postgres smoke test runs when TEST_POSTGRES_URI is available.

Surface area

• Frontend UI — page / component / setting / interaction under frontend/
• Backend API — endpoint / SSE event / request-response shape under backend/app
• Agents / LangGraph — agent node, graph wiring, langgraph.json, or prompt change
• Sandbox — docker/ or sandboxed execution
• Skills — change under skills/
• Dependencies — new/upgraded entry in backend/pyproject.toml or frontend/package.json (say what it buys us)
• Default behavior change — changes existing behavior without the user opting in (default model, default setting, data shape)
• Docs / tests / CI only — no runtime behavior change

Screenshots / Recording

Not applicable; this changes a benchmark and its JSONL output.

Validation

• pytest tests/test_bench_checkpoint_channels.py tests/test_summarize_checkpoint_channels.py tests/test_delta_channel_checkpointers.py -q — 48 passed, 5 skipped
• ruff check . — passed
• ruff format --check . — passed (1,234 files)
• Controller smoke: memory + SQLite, full + delta, 4 child-process rows, schema v2, failures=0
• Real Postgres smoke was not run locally because TEST_POSTGRES_URI is not configured; the conditional test remains skipped in that environment.
• Full offline suite (PowerShell: $env:PYTHONPATH='.;packages/harness;packages/extension-api'; uv run pytest -m "not live" tests/ -q) — 11,993 passed, 128 skipped, 230 failed, 82 errors in 20:17. None of the failures/errors are in the three related test modules; the non-zero cases are retained as current Windows/sparse-checkout environment evidence.

AI assistance

Tool(s) used: Codex

How you used it: Used for issue and repository analysis, implementation assistance, test design, documentation, and validation.

• I've read and understand every line of this change and take responsibility for it — it's not unreviewed AI output.