chore(fraise): move to server v0.3.2

[RonsenbergVI]

Pull Request

What does this PR do?

Moves fraise to server v0.3.2, pinned by digest (sha256:5796b4d5…, amd64 + arm64, commit d123199b). It skips 0.3.1. Both are patch releases on the 0.3 line, so the SDK pin (fraise-sdk>=0.1.1,<0.2, verified against server >=0.3.0,<0.4.0) and the adapter stay as they are, and uv.lock is unchanged. AGENTS.md's pin list is updated to match.

Upstream changes that matter here:

• Recall behaviour changes (#524). Window scoring (db.window-gamma, default 0.3) is on by default and scores a fact together with the turns written just before and after it. Facet spread (the db.aggregate.name default) also spreads top hits across topics. The adapter sets neither, so the defaults apply.
• Decay half-life default goes from 1 week to 90 days (#513). The adapter doesn't set it.
• num-graphs is now capped at 256 at startup (#498). The compose file's -num-graphs 256 is exactly at the cap, which is still accepted.
• The rest of 0.3.1 and 0.3.2 is fixes: query-grammar repairs, rp-tree pool sizing and determinism, delete cleanup, and config validation.

Kind of change

• New memory system integration (benchmarks/<name>/)
• Fix to an existing integration
• New dataset / dataset loader change
• New metric / evaluation axis
• Harness (CLI, reporting, plotting, CI)
• Docs only

Testing

Tested against a local server started from the pinned digest:

docker run -d -p 127.0.0.1:19878:9876 ghcr.io/fraisehq/fraise@sha256:5796b4d589b7572428fe04fc683128baa7c2561c9fa08f62bda62bc7dd74fd1f -num-graphs 256
uv sync --package fraise

server_version() returns 0.3.2, check_compatibility() returns True, and stats() reports 256 graphs.

Keyless (no tagger, no embedder, no OpenAI calls). This uses the same scope as #210, which got 0.913 on 0.3.0:

FRAISE_BASE_URL=http://127.0.0.1:19878 uv run amb run --system fraise --dataset locomo \
  --limit 1 --turns 200 --questions 25 --param model=none --param embedding_model=none

"system_version": "0.3.2",
"num_questions": 25,
"retrieval_precision": 0.3812539682539683,
"retrieval_recall": 0.9133333333333334,
"errors": 0,

No keyed run and no agentic run.

• uv run pytest passes (386 tests). The pre-commit hooks pass too.
• Ran the change against real data, not just unit tests (commands and output snippets are above)

Results / report hygiene

• This PR does not hand-edit runs/, plots/, or RESULTS.md. Those are only ever
  written by amb report / amb plot all, and only committed by the release CI workflow,
  never by hand (CI enforces this). README.md is hand-written — its results section embeds
  images from plots/ by stable path and needs no machine edits
• If this changes what a metric measures, existing numbers in RESULTS.md may now be stale —
  noted below, or flagged for a re-run

system_version becomes 0.3.2, so the next fraise run adds a new results row beside the 0.3.0 one and does not replace it. The smoke number above comes from a small run and can't be compared with published results.