brain-of-reese/tests/integration/test_chunks_hnsw_recall.py

"""Integration: the cosine tier USES ``ix_chunks_embedding_hnsw`` and the
top-100 did not move (phase 188, task 03 — the deterministic half of D4).

**The production numbers this pin stands in for** (read 2026-10-15, read-only):
414,220 chunks — all with a non-null ``embedding`` — a ``chunks`` relation of
≈ **3,000 MB** (database 3,147 MB), ``shared_buffers=128MB``, ``work_mem=4MB``,
a 3,904 MB host with **no swap**, pgvector **0.8.0**, and ``pg_indexes`` for
``chunks`` listing only ``chunks_pkey``, ``ix_chunks_document_id`` and
``ix_chunks_tsv`` (GIN): **no index on ``embedding``**. Every chat turn's A7
cosine tier (``ORDER BY embedding <=> :vec LIMIT 100``) was therefore a
sequential scan of that 3 GB relation.

Two CI pins, both deterministic and LLM-free:

1. **The plan pin** — ``EXPLAIN (FORMAT JSON)`` of the tier's own SELECT (the
   statement object captured from
   :func:`app.rag.retriever._vector_candidates`, compiled with literal binds so
   the planner sees a real constant) names ``ix_chunks_embedding_hnsw`` and
   contains no ``Seq Scan on chunks``. Its vacuity guard runs the SAME
   statement with index scans refused and MUST see the seq scan, so the pin
   cannot pass on a plan the index never influenced.
2. **The recall pin** — the index-backed top-100 against the EXACT cosine
   top-100 of the fixture corpus: ``|overlap| / 100 >= 0.99`` for EVERY query
   at the SHIPPED ``ef_search`` (400 since task 04's live battery; 100, the
   A7 candidate count, was the task-02 default and the battery retired it),
   each returned row's ``1 − distance`` equal to the exact cosine within
   **1e-6**, and the mean one step DEEPER never below the mean at the shipped
   value. Every recall run also re-reads the plan it just ran, because an
   ``ef_search`` the planner declines turns the "index" pin into a second
   measurement of the sequential scan — a vacuous pass, and the one failure
   mode this file exists to catch.

**Measured on this corpus** (5,000 chunks, ``TOP_K = 100``, 30 fixed-seed
queries, ``m = 16`` / ``ef_construction = 64`` from migration 0029):
``ef_search=100`` → mean top-100 overlap **0.9993**, per-query minimum **0.99**
(two of the thirty return 99/100), zero queries below the floor;
``ef_search=200`` and the shipped **400** → mean **1.0000**. Worst
``|cosine − exact cosine|`` over every
returned row: **5.7e-07**, and 1e-6 is the tightest tolerance that can hold —
``_vector_candidates`` publishes ``round(1 − distance, 6)``, which alone can move
a value by 5e-7, on top of the float4 rounding of the distance itself (~6e-8 at
this magnitude). The exact reference is computed over the vectors as the
database returned them, so the comparison measures the index. Setup stays inside the runtime guard:
generate ≈ 0.5 s, seed ≈ 2 s, ``CREATE INDEX CONCURRENTLY`` ≈ 1.4 s, read-back +
exact top-100 ≈ 3.5 s — ≈ 8 s against the ~60 s target.

**What 5,000 rows proves and what it does not.** It is enough for a PLAN pin
(the planner needs a populated relation to prefer an ordered index scan over a
seq scan + top-N sort) and enough for a deterministic index-vs-exact
regression; it is not a proxy for 414k rows, because recall is a function of the
corpus shape and of graph density at 83× the size. The prod-scale gate is task
04 — the live retrieval battery, index off vs on, over the 1,000-document
replica KB, where every ``top_score``, top-5 set and A8 verdict must match.

**Corpus shape, and why it is not uniform noise.** One
``random.Random(20261015)`` stream in a fixed order: a 16-direction principal
subspace (one random unit direction per disjoint 48-dim block of the 768 —
disjoint blocks are orthonormal by construction, so no Gram–Schmidt and no
numerical drift), 40 unit centroids inside it, 5,000 chunks of
``normalise(centroid + 2.0 × a random subspace direction)`` (125 per cluster),
then 30 queries drawn uniformly from the same subspace. Low EFFECTIVE rank (16
of the 768 stored dimensions carry the structure; every dimension is still
non-zero) is deliberate: it is how trained text embeddings behave (their
spectrum concentrates far below the nominal dimension), and it is the shape
under which ``ef_search == LIMIT`` can be expected to return the exact top-100.

The counter-case is recorded here because measuring it is a finding of this
task: the same 5,000 rows with the cluster bias thinned to a whisper (centroid
weight 0.1 over per-dimension Gaussian noise across all 768 dimensions —
effectively iid, the hardest shape a graph index can be handed, and the one
this file's sweep ran at centroid weights 0.1, 0.3 and 0.5 alike) measures mean
recall@100 ≈ **0.63** at ``ef_search=100``, ≈ **0.85** at 200 and **1.00** from
400 up, whatever ``m`` / ``ef_construction`` migration 0029 ships. ``ef_search``
at the candidate count is therefore NOT a free win on high-effective-rank data;
if the live battery moves a ``top_score`` or a top-5 set, LOCKED A3's answer is
to raise ``BOR_RETRIEVAL_EF_SEARCH`` (towards 400), never to relax an assertion
here.

The index is built by migration **0029** and never by an inline ``CREATE
INDEX`` — the migration is the single definition of the index these pins name
(``0029 → 0028`` → seed unindexed → ``upgrade 0029``). As in
``test_chunks_hnsw_index.py``, the seeding session holds NO open transaction
while the ``CONCURRENTLY`` DDL runs (an idle-in-transaction snapshot parks the
build), and teardown leaves the dev DB at head, index valid, fixture rows gone.
"""
from __future__ import annotations

import math
import operator
import random
import struct
import uuid
from collections.abc import Iterator
from typing import Any

import pytest
from alembic.config import Config
from sqlalchemy import select, text
from sqlalchemy.dialects import postgresql
from sqlalchemy.orm import Session

from alembic import command
from app.config import Settings
from app.db import SessionLocal, db_available
from app.models import Chunk, Document
from app.rag import retriever
from app.rag.retriever import _vector_candidates

# ---------------------------------------------------------------------------
# The corpus (D4's deterministic half) — one fixed stream, no wall clock
# ---------------------------------------------------------------------------

DIM = 768  # A6: the embedding dimension is fixed at table creation
CENTROIDS = 40
SUBSPACE = 16  # effective rank of the corpus (see the module docstring)
BLOCK = DIM // SUBSPACE  # 48 dims per basis direction, disjoint by construction
NOISE = 2.0  # |noise| / |centroid| — cos(chunk, own centroid) ~= 0.45
CORPUS_CHUNKS = 5000
CHUNKS_PER_DOCUMENT = 10
QUERIES = 30
SEED = 20261015  # the date of the incident that started phase 188
FIXTURE_SOURCE = "phase-188-hnsw-recall"

TOP_K = 100  # A7's cosine candidate count — the LIMIT the tier promises
EF_PIN = 400  # the shipped default (Settings.retrieval_ef_search)
EF_RAISED = 600  # the monotonicity probe
RECALL_FLOOR = 0.99  # LOCKED A3 — a raised ef_search is the only remedy
COSINE_TOLERANCE = 1e-6  # "exact to 6 decimals"

#: The pins need a small relation: the exact reference is a Python scan over
#: every row. A DB holding more is a restored replica rather than the dev
#: corpus — that scale is task 04's live battery, and 30 × N dot products here
#: would only make the suite crawl.
MAX_RELATION_ROWS = 20_000

INDEX_NAME = "ix_chunks_embedding_hnsw"
RELATION = "chunks"


def _f32(value: float) -> float:
    """The ``float4`` the database holds for a query component.

    A question vector goes to Postgres as ``str(float)`` text and comes back as
    ``float4`` (``strtof`` rounds to nearest, the same value ``struct.pack``
    produces), so the exact reference rounds the QUERY the way the server does.
    The corpus needs no modelling — its components are read back from the
    database — which is what makes the comparison measure the INDEX rather than
    a precision artefact of the fixture.
    """
    return struct.unpack("f", struct.pack("f", value))[0]


def _norm(vec: list[float]) -> float:
    return math.sqrt(sum(v * v for v in vec))


def _unit(vec: list[float]) -> list[float]:
    n = _norm(vec)
    return [v / n for v in vec] if n else vec


def _generate() -> tuple[list[list[float]], list[list[float]]]:
    """Fixed-seed ``(corpus, queries)``, all L2-normalised.

    Normalising makes cosine the dot product, which is what keeps the exact
    reference cheap enough to run in CI. One stream in a fixed order (basis →
    centroids → chunks → queries) is what turns the recall floor into a
    regression pin instead of a measurement: byte-identical input, therefore a
    byte-identical graph, therefore a byte-identical overlap.
    """
    rng = random.Random(SEED)
    basis = [_unit([rng.gauss(0.0, 1.0) for _ in range(BLOCK)]) for _ in range(SUBSPACE)]

    def subspace_vector() -> list[float]:
        """A random unit vector inside the 16-direction subspace."""
        coefs = [rng.gauss(0.0, 1.0) for _ in range(SUBSPACE)]
        n = _norm(coefs) or 1.0
        out = [0.0] * DIM
        for k, direction in enumerate(basis):
            weight = coefs[k] / n
            start = k * BLOCK
            for j in range(BLOCK):
                out[start + j] = weight * direction[j]
        return out

    centroids = [subspace_vector() for _ in range(CENTROIDS)]
    corpus: list[list[float]] = []
    for i in range(CORPUS_CHUNKS):
        centroid = centroids[i % CENTROIDS]
        noise = subspace_vector()
        corpus.append(_unit([c + NOISE * n for c, n in zip(centroid, noise, strict=True)]))
    queries = [subspace_vector() for _ in range(QUERIES)]
    return corpus, queries


def _seed(session: Session, corpus: list[list[float]]) -> None:
    """Insert the corpus — 500 fixture documents × 10 chunks, one commit.

    Ordinary content chunks (``is_summary`` False, embedding set): the
    NULL-embedding skip is pinned in ``tests/unit/test_retriever.py``, and a
    NULL row would sit in neither the exact set nor the index.
    """
    documents = [
        Document(
            id=uuid.uuid4(),
            source=FIXTURE_SOURCE,
            path=f"{FIXTURE_SOURCE}/{d}.md",
            full_path=f"/tmp/{FIXTURE_SOURCE}/{d}.md",
            title=f"hnsw recall fixture {d}",
            content="fixture content",
            content_hash="0" * 64,
        )
        for d in range(CORPUS_CHUNKS // CHUNKS_PER_DOCUMENT)
    ]
    session.add_all(documents)
    session.flush()
    session.add_all(
        [
            Chunk(
                id=uuid.uuid4(),
                document_id=documents[i // CHUNKS_PER_DOCUMENT].id,
                position=i % CHUNKS_PER_DOCUMENT,
                content=f"fixture chunk {i}",
                embedding=vector,
                is_summary=False,
            )
            for i, vector in enumerate(corpus)
        ]
    )
    session.commit()


# ---------------------------------------------------------------------------
# Exact reference (pure Python, float4-modelled inputs)
# ---------------------------------------------------------------------------


def _exact_top(stored: dict[Any, list[float]], query: list[float]) -> list[tuple[float, Any]]:
    """The EXACT cosine top-:data:`TOP_K` of *stored* for *query*.

    Mirrors pgvector's ``vector_cosine_distance``: three float8 accumulators
    (``dot``, ``norm_a``, ``norm_b``) over the components left to right, then
    ``dot / sqrt(norm_a * norm_b)``. The norms are recomputed from the stored
    components rather than assumed to be 1, because the stored component is the
    float4 of a normalised double.

    Pure Python (no numpy in this project — A3 keeps the dependency list
    closed): ``sum(map(mul, …))`` is ~25 µs per 768-dim dot product, so 30
    queries over 5,000 rows cost ~3 s. The one residual inaccuracy is the
    read-back itself: pgvector prints each float4 in shortest-round-trip form,
    so a component comes back within ~1 ulp of the float4 (measured max
    7.4e-09 absolute over the fixture corpus) — three orders below the 1e-6
    this file pins.
    """
    q = [_f32(v) for v in query]
    q_norm = math.sqrt(sum(map(operator.mul, q, q)))
    scored = [
        (
            sum(map(operator.mul, vec, q)) / (math.sqrt(sum(map(operator.mul, vec, vec))) * q_norm),
            chunk_id,
        )
        for chunk_id, vec in stored.items()
    ]
    scored.sort(key=lambda item: -item[0])
    return scored[:TOP_K]


# ---------------------------------------------------------------------------
# Capturing the tier's own statement (the plan pin must explain THAT SQL)
# ---------------------------------------------------------------------------


class _StatementRecorder:
    """Records what :func:`_vector_candidates` sends, answering with no rows."""

    def __init__(self) -> None:
        self.statements: list[Any] = []

    def execute(self, statement: Any, params: Any = None) -> _StatementRecorder:
        self.statements.append(statement)
        return self

    def all(self) -> list[Any]:
        return []


def _vector_statement_sql() -> str:
    """The tier's cosine SELECT as one literal SQL string, ready for EXPLAIN.

    A hand-written copy would drift the moment the retriever changes and would
    then pin SQL no request ever runs, so the statement OBJECT is captured.
    ``literal_binds`` gives the planner a real vector constant
    (``pgvector.sqlalchemy.VECTOR`` ships a literal processor, and the resulting
    unknown string literal coerces to ``vector`` through the ``<=>`` operator's
    own signature). The ``SET LOCAL`` line is deliberately not part of it:
    ``EXPLAIN`` takes one statement, and that line tunes the search depth, not
    the shape of the plan.
    """
    recorder = _StatementRecorder()
    _vector_candidates(recorder, [0.0] * DIM, TOP_K)  # pyright: ignore[reportArgumentType]
    set_local, select_statement = recorder.statements
    assert str(set_local) == f"SET LOCAL hnsw.ef_search = {EF_PIN}", (
        f"the tier's ef_search line changed ({set_local}) — the capture must be the SELECT"
    )
    sql = str(
        select_statement.compile(
            dialect=postgresql.dialect(), compile_kwargs={"literal_binds": True}
        )
    )
    assert "<=>" in sql, f"the cosine operator is gone from the tier: {sql[:200]}"
    return sql


def _plan(session: Session, sql: str) -> list[dict[str, Any]]:
    """Every node of ``EXPLAIN (FORMAT JSON)`` for *sql*, flattened."""
    document = session.execute(text(f"EXPLAIN (FORMAT JSON) {sql}")).scalar()
    assert isinstance(document, list) and document, f"no plan for {sql[:120]}: {document!r}"
    nodes: list[dict[str, Any]] = []

    def walk(node: dict[str, Any]) -> None:
        nodes.append(node)
        for child in node.get("Plans", []):
            walk(child)

    walk(document[0]["Plan"])
    return nodes


def _describe(nodes: list[dict[str, Any]]) -> str:
    return ", ".join(
        f"{node.get('Node Type')}[{node.get('Index Name') or node.get('Relation Name') or '?'}]"
        for node in nodes
    )


def _uses_index(nodes: list[dict[str, Any]], index_name: str) -> bool:
    return any(
        "Index Scan" in str(node.get("Node Type", "")) and node.get("Index Name") == index_name
        for node in nodes
    )


def _seq_scans(nodes: list[dict[str, Any]], relation: str) -> bool:
    return any(
        node.get("Node Type") == "Seq Scan" and node.get("Relation Name") == relation
        for node in nodes
    )


# ---------------------------------------------------------------------------
# Fixture: seed unindexed, build through 0029, read the corpus back once
# ---------------------------------------------------------------------------


@pytest.fixture(scope="module")
def hnsw_corpus() -> Iterator[dict[str, Any]]:
    """Seed the corpus, build the index with migration 0029, precompute exact.

    ``0029 → 0028`` → seed → ``0029``: seeding with the index present would pay
    a graph insert per row, and building after the load is what the operator
    does on the production box anyway. Teardown deletes ONLY this module's rows
    (``source = FIXTURE_SOURCE``; the chunk rows go with them by the DB-level
    cascade — AGENTS rule 12, no chunk hydration) and leaves the dev DB at head
    with the index valid, exactly as it was found.
    """
    if not db_available():
        pytest.skip("Postgres not reachable — run `podman compose up -d db` first")
    session = SessionLocal()
    config = Config()  # no alembic.ini — env.py reads the URL from app config
    config.set_main_option("script_location", "alembic")
    try:
        preexisting = session.execute(text(f"SELECT count(*) FROM {RELATION}")).scalar() or 0
        if preexisting > MAX_RELATION_ROWS:
            pytest.skip(
                f"{RELATION} holds {preexisting} rows: the exact reference is a Python "
                "scan, so these pins need the small dev corpus. Production scale "
                "is task 04's live battery."
            )
        corpus, queries = _generate()
        session.rollback()  # no open snapshot while the CONCURRENTLY DDL runs
        command.downgrade(config, "0028")
        session.rollback()
        _seed(session, corpus)
        session.rollback()
        command.upgrade(config, "0029")  # the ONE definition of the index
        session.rollback()
        # Planner statistics for the plan pin: a freshly loaded relation has
        # none, and a row estimate of 0 makes "seq scan + top-N sort" look free.
        session.execute(text(f"ANALYZE {RELATION}"))
        session.commit()
        rows = session.execute(select(Chunk.id, Chunk.embedding)).all()
        stored: dict[Any, list[float]] = {row[0]: row[1] or [] for row in rows}
        assert len(stored) >= preexisting + CORPUS_CHUNKS, (
            f"expected {CORPUS_CHUNKS} fixture rows on top of {preexisting}, "
            f"the relation holds {len(stored)}"
        )
        # The exact sets are the expensive half (30 × N dot products); the two
        # recall pins differ only in the ef_search they ask the index for.
        yield {
            "queries": queries,
            "exact": [_exact_top(stored, query) for query in queries],
            "chunk_ids": set(stored),
            "plan_sql": _vector_statement_sql(),
        }
    finally:
        session.rollback()
        session.execute(
            text("DELETE FROM documents WHERE source = :src"), {"src": FIXTURE_SOURCE}
        )
        session.commit()
        session.rollback()
        command.upgrade(config, "head")  # the dev DB never stays below head
        session.close()


def _stub_settings(**kwargs: object) -> Settings:
    """A :class:`Settings` off the env file (the ``test_chunks_hnsw_index.py``
    helper's shape — an operator's shell must not tune these pins)."""
    return Settings(_env_file=None, **kwargs)  # pyright: ignore[reportCallIssue]


def _force_index_scan(session: Session) -> None:
    """Put the planner on the index, because a recall pin has to measure the
    index and not the planner's preference.

    Postgres costs an HNSW scan partly on ``ef_search``, so on a relation this
    small it can legitimately prefer a seq scan + top-N sort — and a recall run
    that did that would score the EXACT scan twice and report a perfect 1.0000
    without visiting a single graph node. That is not hypothetical: the live
    gate (``benchmarks/phase188_live_gate/``) watched exactly this happen on the
    8,866-chunk replica, where the planner used the index at ``ef_search = 40``
    and refused it at 100.

    ``enable_seqscan = off`` is Postgres' own "take any other plan if one
    exists" lever. It is set at SESSION scope because ``_vector_candidates``
    opens its own transaction, so a ``SET LOCAL`` issued here would die with the
    caller's transaction instead of reaching the tier. Every caller resets it in
    a ``finally`` — a leaked ``off`` would slow every later test on this pooled
    connection — and the assertion beside it proves the lever took effect.
    """
    session.rollback()
    session.execute(text("SET enable_seqscan = off"))


def _release_index_scan(session: Session) -> None:
    session.rollback()
    session.execute(text("RESET enable_seqscan"))
    session.rollback()


def _set_ef_search(monkeypatch: pytest.MonkeyPatch, value: int) -> None:
    """Drive the tier at *value* the way an operator does — the setting."""
    settings = _stub_settings(retrieval_ef_search=value)
    monkeypatch.setattr(retriever, "get_settings", lambda: settings)


# ---------------------------------------------------------------------------
# Pin 1 — the planner uses the index (with its vacuity guard)
# ---------------------------------------------------------------------------


def test_plan_uses_the_hnsw_index(db: Session, hnsw_corpus: dict[str, Any]) -> None:
    """The 2026-10-15 incident WAS a plan, so the pin is a plan: the tier's own
    ``ORDER BY embedding <=> <vector> LIMIT 100`` runs through
    ``ix_chunks_embedding_hnsw`` and never reads ``chunks`` sequentially.

    Explaining the captured statement is also what proves D1's "no query
    rewrite": the ``<=>`` the retriever already emits is exactly the operator
    ``vector_cosine_ops`` serves.
    """
    nodes = _plan(db, hnsw_corpus["plan_sql"])
    assert _uses_index(nodes, INDEX_NAME), (
        f"{INDEX_NAME} is not in the plan — the incident is back: [{_describe(nodes)}]"
    )
    assert not _seq_scans(nodes, RELATION), (
        f"a sequential scan of {RELATION} survived the index: [{_describe(nodes)}]"
    )
    db.rollback()


def test_plan_without_the_index_seq_scans_the_relation(
    db: Session, hnsw_corpus: dict[str, Any]
) -> None:
    """Vacuity guard: the same statement, same session, same statistics, with
    index scans refused, MUST fall back to ``Seq Scan on chunks``.

    ``SET LOCAL enable_indexscan = off`` rather than dropping the index: a
    ``DROP INDEX CONCURRENTLY`` + rebuild would cost a second graph build and
    leave the corpus unindexed if the rebuild failed, while this GUC is
    transaction-scoped (it dies with the rollback) and removes exactly the
    access path under test. Without this test the plan pin would pass vacuously
    on any plan that merely mentions some index.
    """
    sql = hnsw_corpus["plan_sql"]
    db.rollback()
    db.execute(text("SET LOCAL enable_indexscan = off"))
    db.execute(text("SET LOCAL enable_indexonlyscan = off"))
    nodes = _plan(db, sql)
    assert _seq_scans(nodes, RELATION), (
        f"the vacuity guard itself did not seq-scan {RELATION}: [{_describe(nodes)}]"
    )
    assert not _uses_index(nodes, INDEX_NAME), (
        f"{INDEX_NAME} was used with index scans disabled: [{_describe(nodes)}]"
    )
    db.rollback()


# ---------------------------------------------------------------------------
# Pin 2 — the approximate top-100 IS the exact top-100
# ---------------------------------------------------------------------------


def test_top100_recall_at_the_shipped_ef_search(
    db: Session,
    hnsw_corpus: dict[str, Any],
    monkeypatch: pytest.MonkeyPatch,
) -> None:
    """At the shipped ``ef_search`` the index returns the exact top-100 for
    every query, exact to 1e-6 on the score the honesty gate reads.

    Two different failures surface here: a ``top_score`` that moved — A8's
    deflect/answer verdict flips on that one number, so it is pinned per row,
    not per query — and an overlap below the floor, i.e. a candidate that never
    reaches RRF. The failure message carries every query's overlap so a future
    ``ef_search`` raise has its evidence in the report instead of in a re-run.
    """
    queries: list[list[float]] = hnsw_corpus["queries"]
    exact: list[list[tuple[float, Any]]] = hnsw_corpus["exact"]
    chunk_ids: set[Any] = hnsw_corpus["chunk_ids"]
    _set_ef_search(monkeypatch, EF_PIN)

    _force_index_scan(db)
    try:
        depth_nodes = _plan(db, hnsw_corpus["plan_sql"])
        assert _uses_index(depth_nodes, INDEX_NAME), (
            f"this recall run at ef_search={EF_PIN} did not use {INDEX_NAME} even "
            "with sequential scans refused — the pin would be measuring a scan "
            f"and calling it parity: [{_describe(depth_nodes)}]"
        )

        overlaps: list[float] = []
        worst_delta = 0.0
        for index, (query, exact_rows) in enumerate(zip(queries, exact, strict=True)):
            exact_ids = {chunk_id for _, chunk_id in exact_rows}
            exact_scores = {chunk_id: cosine for cosine, chunk_id in exact_rows}
            candidates = _vector_candidates(db, query, TOP_K)
            db.rollback()  # end the turn's transaction, SET LOCAL included
            assert len(candidates) == TOP_K, (
                f"query {index}: the tier returned {len(candidates)} candidates, "
                f"not the {TOP_K} the A7 cosine tier promises"
            )
            got_ids = {candidate.chunk_id for candidate in candidates}
            assert got_ids <= chunk_ids, (
                f"query {index}: a candidate is outside the corpus the exact "
                "reference was computed over"
            )
            overlaps.append(len(exact_ids & got_ids) / TOP_K)
            for candidate in candidates:
                exact_cosine = exact_scores.get(candidate.chunk_id)
                if exact_cosine is not None:
                    worst_delta = max(worst_delta, abs(candidate.cosine - exact_cosine))
    finally:
        _release_index_scan(db)

    mean = sum(overlaps) / len(overlaps)
    below = [(i, overlap) for i, overlap in enumerate(overlaps) if overlap < RECALL_FLOOR]
    assert not below, (
        f"top-{TOP_K} recall below {RECALL_FLOOR} at ef_search={EF_PIN} for "
        f"{len(below)}/{len(overlaps)} queries: {below}; per-query overlap="
        f"{[round(overlap, 4) for overlap in overlaps]} mean={mean:.4f} — the "
        "remedy is a higher BOR_RETRIEVAL_EF_SEARCH, never this floor (LOCKED A3)"
    )
    assert worst_delta < COSINE_TOLERANCE, (
        f"1 - distance drifted from the exact cosine by {worst_delta:.3e} "
        f"(tolerance {COSINE_TOLERANCE:.0e}) — top_score would not be exact"
    )


def test_recall_never_below_the_pinned_floor_when_ef_search_raised(
    db: Session,
    hnsw_corpus: dict[str, Any],
    monkeypatch: pytest.MonkeyPatch,
) -> None:
    """Monotonicity sanity: a deeper search cannot recover FEWER of the exact
    top-100 than a shallower one.

    This is the property the operator leans on when the live battery (task 04)
    says "raise it" — ``ef_search`` is a recall dial, not a lottery. A
    regression here means the value never reaches the server (the ``SET LOCAL``
    of task 02), which is exactly what the shipped 400 would then silently not
    do. ``EF_RAISED`` sits ABOVE the shipped value, so the probe still measures
    a deeper search after task 04 moved the default.
    """
    exact: list[list[tuple[float, Any]]] = hnsw_corpus["exact"]
    queries: list[list[float]] = hnsw_corpus["queries"]

    def mean_overlap(ef_search: int) -> float:
        _set_ef_search(monkeypatch, ef_search)
        total = 0.0
        for query, exact_rows in zip(queries, exact, strict=True):
            exact_ids = {chunk_id for _, chunk_id in exact_rows}
            candidates = _vector_candidates(db, query, TOP_K)
            db.rollback()
            total += len(exact_ids & {candidate.chunk_id for candidate in candidates}) / TOP_K
        return total / len(exact)

    pinned = mean_overlap(EF_PIN)
    raised = mean_overlap(EF_RAISED)
    assert raised >= pinned, (
        f"ef_search={EF_RAISED} recovered less of the exact top-{TOP_K} "
        f"({raised:.4f}) than ef_search={EF_PIN} ({pinned:.4f})"
    )