release: v1.108.185 — fusion gets a verdict, and fusion can prove abs… #137

	name: Replay

	# Retrieval-quality regression gate. Indexes this repo, replays a fixed set of
	# golden queries through search_symbols, and fails if nDCG / MRR / Recall@k drop
	# more than 2% (relative) below the committed golden baseline. This is the gate
	# that lets ranking-affecting changes (fusion weights, BM25 normalization, parser
	# extraction) land with a proof they did not regress retrieval quality.

	on:
	push:
	branches: [main]
	pull_request:

	permissions:
	contents: read

	jobs:
	replay:
	name: Retrieval-quality gate
	runs-on: ubuntu-latest
	steps:
	- uses: actions/checkout@93cb6efe18208431cddfb8368fd83d5badbf9bfd # v5
	- name: Set up uv
	uses: astral-sh/setup-uv@37802adc94f370d6bfd71619e3f0bf239e1f3b78 # v7
	with:
	python-version: "3.12"
	- name: Install
	run: uv sync --group dev
	- name: Index self + run replay gate
	run: \|
	set -euo pipefail
	STORE="${RUNNER_TEMP}/replay-index"
	# The repo id is a hash of the absolute index path, which differs on a
	# CI runner, so capture the actual id and pass it through --repo.
	REPO_ID=$(uv run python -c "from jcodemunch_mcp.tools.index_folder import index_folder; print(index_folder(path='.', use_ai_summaries=False, incremental=False, storage_path='${STORE}')['repo'])")
	echo "indexed repo: ${REPO_ID}"
	uv run python benchmarks/replay/run_replay.py \
	--fixture benchmarks/replay/fixtures/self_v1_75_0.json \
	--repo "${REPO_ID}" \
	--storage-path "${STORE}" \
	--baseline-file benchmarks/replay/results/self_v1_75_0-golden.json \
	--gate 0.02

Provide feedback