6krrt/tests/test_routing_profiles_integration.py

"""Integration pins for the five preserved behaviors of named routing profiles.

This file covers Todo 7 of the named-routing-profiles plan. These tests cut
across the dispatcher / routing boundary: `route` is invoked directly with a
real routing profile, against a monkeypatched SQLite catalog in a temp file.
No classifier, provider, or local model is ever called.

The five planks:

(a) ``auto`` and ``auto:batch`` behavior is byte-identical to the old ternary:
    route() with the batch profile must return the same selection and
    runner-up ordering as the default profile with request
    ``latency_tolerance=batch``.
(b) Unknown profiles return 422 naming valid profiles. Coverage for
    ``/v1/chat/completions`` unknown profiles exists in
    tests/test_chat_completions.py; we do not duplicate it here.
(c) Empty profile allowlist fails with a profile-specific reason
    ("profile_excluded") and the reason is persisted to route_decisions
    when no candidate survives.
(d) Profiles AND with circuit-breaker exclusions, admin overrides, and
    ``eligible_categories``.
(e) Exploration picks only from profile-filtered candidates.
"""

from __future__ import annotations

import json
import random
import sqlite3
from pathlib import Path
from typing import Optional

import pytest
from starlette.testclient import TestClient

import circuit_breaker
import dispatcher
from config import RoutingProfile
from dispatcher import (
    BUILTIN_PROFILES,
    TaskRequest,
    app,
    persist_route_decision,
    route,
)
from routing import rejection_reason, select_candidates

ROOT = Path(__file__).resolve().parent.parent
SCHEMA_SQL = (ROOT / "config" / "schema.sql").read_text()
ADMIN_SCHEMA_SQL = (ROOT / "config" / "admin_schema.sql").read_text()

M1 = "m1-cloud"
M1_FLEX = "m1-cloud-flex"
M2 = "m2-cloud"
LOCAL = "m3-local"
LOCAL2 = "m3b-local"


def _model_row(
    model_id: str,
    *,
    provider: str = "neuralwatt",
    tier: int = 2,
    latency_class: str = "standard",
    completion_price: float = 1.0,
    prompt_price: Optional[float] = None,
    context_window: int = 262128,
    effective_context_window: int = 192500,
    supports_vision: int = 0,
    supports_json_mode: int = 1,
    eligible_categories: Optional[str] = None,
) -> dict:
    """A models-table row, shaped for the fixture's INSERT."""
    return {
        "model_id": model_id,
        "provider": provider,
        "base_model_id": model_id,
        "tier": tier,
        "context_window": context_window,
        "effective_context_window": effective_context_window,
        "max_output_tokens": 16384,
        "cost_per_1m_prompt": (
            prompt_price if prompt_price is not None else completion_price / 3
        ),
        "cost_per_1m_completion": completion_price,
        "cost_per_1m_prompt_cached": (
            prompt_price if prompt_price is not None else completion_price / 3
        ),
        "supports_vision": supports_vision,
        "supports_json_mode": supports_json_mode,
        "latency_class": latency_class,
        "reasoning_mode": "default",
        "context_variant": "full",
        "access_level": "public",
        "availability": "active",
        "eligible_categories": eligible_categories,
        "last_updated": "2026-08-22T00:00:00+00:00",
    }


def _insert_model(conn: sqlite3.Connection, row: dict) -> None:
    conn.execute(
        """
        INSERT INTO models (
            model_id, provider, base_model_id, tier, context_window,
            effective_context_window, max_output_tokens,
            cost_per_1m_prompt, cost_per_1m_completion, cost_per_1m_prompt_cached,
            supports_vision, supports_json_mode,
            latency_class, reasoning_mode, context_variant,
            access_level, availability, eligible_categories, last_updated
        ) VALUES (
            :model_id, :provider, :base_model_id, :tier, :context_window,
            :effective_context_window, :max_output_tokens,
            :cost_per_1m_prompt, :cost_per_1m_completion, :cost_per_1m_prompt_cached,
            :supports_vision, :supports_json_mode,
            :latency_class, :reasoning_mode, :context_variant,
            :access_level, :availability, :eligible_categories, :last_updated
        )
        """,
        row,
    )


def _insert_proficiency(
    conn: sqlite3.Connection,
    model_id: str,
    provider: str,
    *,
    blended_score: float,
    outcome_samples: int = 0,
) -> None:
    conn.execute(
        """
        INSERT INTO proficiency (
            model_id, provider, category, blended_score, source,
            outcome_samples, last_updated
        ) VALUES (?, ?, 'coding_general', ?, 'leaderboard', ?,
                  '2026-08-22T00:00:00+00:00')
        """,
        (model_id, provider, blended_score, outcome_samples),
    )


class FakeResponse:
    """Just enough of requests.Response for the stubbed provider POST."""

    def __init__(self, payload=None):
        self.status_code = 200
        self._payload = payload or {}
        self.text = json.dumps(self._payload)
        self.headers = {}
        self.closed = False

    def json(self):
        return self._payload

    def iter_lines(self, decode_unicode=False):
        return iter([])

    def close(self):
        self.closed = True


@pytest.fixture
def profile_router(tmp_path, monkeypatch):
    """The dispatcher pointed at a curated throwaway catalog, nothing dialled out.

    Candidate economics are chosen so the routing outcome is fully determined:

    - M1 / M1_FLEX tie on proficiency (0.9); the flex twin is cheaper, so under
      batch latency it outranks standard, and under interactive it is filtered
      out entirely by the latency_class hard filter.
    - In exploration tests M2 has zero outcome samples, so an epsilon=1.0 run
      with the default profile picks it (cloud), while the locality profile's
      candidate set contains only the two ollama-local rows.
    """
    db_path = tmp_path / "profiles.db"
    conn = sqlite3.connect(db_path)
    conn.executescript(SCHEMA_SQL)
    conn.executescript(ADMIN_SCHEMA_SQL)

    _insert_model(conn, _model_row(M1, completion_price=3.0))
    _insert_model(
        conn, _model_row(M1_FLEX, completion_price=2.0, latency_class="flex")
    )
    _insert_model(conn, _model_row(M2, completion_price=9.0))
    _insert_model(
        conn,
        _model_row(
            LOCAL, provider="ollama-local", completion_price=5.0, prompt_price=1.0
        ),
    )
    _insert_model(
        conn,
        _model_row(
            LOCAL2, provider="ollama-local", completion_price=1.0, prompt_price=0.33
        ),
    )

    _insert_proficiency(conn, M1, "neuralwatt", blended_score=0.90, outcome_samples=50)
    _insert_proficiency(conn, M1_FLEX, "neuralwatt", blended_score=0.90, outcome_samples=50)
    _insert_proficiency(conn, M2, "neuralwatt", blended_score=0.85, outcome_samples=0)
    _insert_proficiency(conn, LOCAL, "ollama-local", blended_score=0.88, outcome_samples=50)
    # 0.70 sits a full quality-tolerance band below the 0.88/0.90 rows
    # (0.9 - 0.8 would land there only by float luck: 0.099999... rounds to
    # band 0). Band 1 keeps LOCAL2 eligible-but-ranked-last under the default
    # profile while still being the cheapest local row under locality.
    _insert_proficiency(conn, LOCAL2, "ollama-local", blended_score=0.70, outcome_samples=50)

    conn.commit()
    conn.close()

    monkeypatch.setattr(dispatcher.cfg.database, "path", str(db_path))
    monkeypatch.setattr(dispatcher.cfg.verification, "local_llm_enabled", False)
    monkeypatch.setattr(dispatcher.cfg.local_vision, "enabled", False)
    monkeypatch.setattr(dispatcher.cfg.session_cache, "enabled", False)
    # Exploration is pinned off for determinism; the exploration tests
    # re-enable it explicitly.
    monkeypatch.setattr(dispatcher.cfg.exploration, "enabled", False)
    monkeypatch.setattr(dispatcher.cfg.exploration, "epsilon", 0.0)
    monkeypatch.setenv("NEURALWATT_API_KEY", "test-key")

    calls = []

    def fake_post(url, headers=None, json=None, stream=False, timeout=None):
        calls.append({"url": url, "body": json, "stream": stream})
        return FakeResponse(
            {
                "id": "chatcmpl-test-1",
                "model": json["model"],
                "choices": [
                    {"message": {"role": "assistant", "content": "ok"},
                     "finish_reason": "stop"}
                ],
                "usage": {"prompt_tokens": 31, "completion_tokens": 12},
                "energy": {"energy_kwh": 5e-05},
                "cost": {"request_cost_usd": 4e-04},
            }
        )

    monkeypatch.setattr(dispatcher.requests, "post", fake_post)
    monkeypatch.setattr(
        dispatcher, "classify",
        lambda task, context: dispatcher.Classification(
            task_category="coding_general", task_tier=2,
            required_context_tokens=100, confidence=0.9,
        ),
    )

    yield TestClient(app), db_path, calls

    circuit_breaker.clear()


@pytest.fixture
def logbuf():
    """Capture what would reach the journal, through the real formatter."""
    import io

    import logs

    buf = io.StringIO()
    logs.configure("debug", stream=buf, journald=False)
    yield buf
    for handler in list(logs.log.handlers):
        logs.log.removeHandler(handler)


def _route_req(
    *,
    latency_tolerance: Optional[str] = None,
    task_category: str = "coding_general",
):
    return TaskRequest(
        task="write me a function",
        task_category=task_category,
        task_tier=2,
        required_context_tokens=100,
        latency_tolerance=latency_tolerance,
    )


def _hard_filters(exclude_models: set[str] = frozenset()) -> dict:
    """The hard-filter kwargs every direct rejection_reason/select_candidates
    call in this file shares, matching what route() itself passes."""
    return dict(
        required_context_tokens=100,
        required_tier=2,
        latency_tolerance="interactive",
        allowed_access_levels=["public"],
        exclude_stale=True,
        exclude_deprecated=True,
        exclude_models=exclude_models,
    )


# =============================================================================
# Plank (a): auto and auto:batch behavior is byte-identical to the old ternary.
# =============================================================================


def test_auto_batch_matches_default_with_batch_latency(profile_router):
    """route() under the batch profile must equal the default profile plus
    request latency_tolerance=batch, byte for byte.

    The batch profile's only field is latency_tolerance, and a request-level
    batch tolerance on the default profile reaches the same effective latency;
    both must produce the same selection, runner-up ordering, and candidate
    count. A leftover special-case ternary that treated ``auto:batch``
    differently from an explicit batch request would break this.
    """
    _, _, _ = profile_router

    via_request = route(
        _route_req(latency_tolerance="batch"),
        profile="default",
        profile_obj=BUILTIN_PROFILES["default"],
    )
    via_profile = route(
        _route_req(latency_tolerance="batch"),
        profile="batch",
        profile_obj=BUILTIN_PROFILES["batch"],
    )

    assert via_request.selected is not None
    assert via_profile.selected is not None
    assert via_request.selected.model_id == via_profile.selected.model_id
    assert [r.model_id for r in via_request.runners_up] == [
        r.model_id for r in via_profile.runners_up
    ]
    assert via_request.candidates_considered == via_profile.candidates_considered
    assert via_request.latency_tolerance == via_profile.latency_tolerance == "batch"
    # The flex twin is cheaper at equal proficiency, so batch must admit it and
    # rank it first: this is what makes the batch equivalence observable.
    assert via_profile.selected.model_id == M1_FLEX
    assert via_profile.selected.latency_class == "flex"


def test_interactive_default_excludes_flex_rows_batch_admits_them(profile_router):
    """Without batch tolerance, the flex twin is latency-filtered out; the
    selection therefore differs, proving the flex scenario is live."""
    _, _, _ = profile_router

    interactive = route(
        _route_req(), profile="default", profile_obj=BUILTIN_PROFILES["default"]
    )
    batch = route(
        _route_req(latency_tolerance="batch"),
        profile="batch",
        profile_obj=BUILTIN_PROFILES["batch"],
    )

    assert interactive.selected is not None
    assert batch.selected is not None
    assert interactive.selected.model_id == M1
    assert interactive.selected.latency_class == "standard"
    assert batch.selected.model_id == M1_FLEX
    assert batch.selected.latency_class == "flex"
    # The flex twin only enters the candidate set under batch tolerance.
    assert interactive.candidates_considered < batch.candidates_considered


def test_chat_auto_batch_selects_same_model_as_auto_under_batch_default(
    profile_router, monkeypatch
):
    """``POST /v1/chat/completions`` with model=auto:batch returns the same
    selected model as model=auto when the configured default profile is batch.

    Since Wave A config profiles may not shadow built-in names, the test sets
    ``routing.default_profile`` to a new config profile named ``batch-local``
    whose only override is ``latency_tolerance=batch``. Both ``auto`` and
    ``auto:batch`` must then pick the flex twin through the profile layer.
    """
    client, _, _ = profile_router
    monkeypatch.setattr(dispatcher.cfg.routing, "default_profile", "batch-local")
    monkeypatch.setitem(
        dispatcher.cfg.profiles, "batch-local", RoutingProfile(latency_tolerance="batch")
    )

    auto = client.post(
        "/v1/chat/completions",
        json={"model": "auto", "messages": [{"role": "user", "content": "hi"}]},
    )
    auto_batch = client.post(
        "/v1/chat/completions",
        json={"model": "auto:batch", "messages": [{"role": "user", "content": "hi"}]},
    )

    assert auto.status_code == 200
    assert auto_batch.status_code == 200
    assert auto.headers["X-Router-Model"] == M1_FLEX
    assert auto_batch.headers["X-Router-Model"] == M1_FLEX
    assert auto.json()["model"] == auto_batch.json()["model"] == M1_FLEX


# =============================================================================
# Plank (b): Unknown profile returns 422 naming valid profiles.
# =============================================================================


def test_route_unknown_profile_returns_422_naming_valid_profiles(profile_router):
    """A misspelled bare profile on /route must name the valid choices."""
    client, _, _ = profile_router
    resp = client.post("/route", json={"task": "x", "profile": "nosuchprofile"})
    assert resp.status_code == 422
    detail = resp.json()["detail"]
    assert "nosuchprofile" in detail
    assert "default" in detail
    assert "batch" in detail
    assert "locality" in detail


# ``/v1/chat/completions`` unknown-profile coverage is NOT duplicated here.
# See tests/test_chat_completions.py::test_unknown_profile_returns_422_naming_valid_profiles
# for the auto:<unknown> 422 naming default/batch, added in Todos 3/4.


# =============================================================================
# Plank (c): Empty profile allowlist fails with a profile-specific reason.
# =============================================================================


def test_ghost_allowlist_selects_nothing_and_persists_profile_excluded(profile_router):
    """A profile whose allowed_model_ids intersect no catalog row selects
    nothing, debug-logs ``profile_excluded`` for every candidate, and can
    persist that reason through the persist_route_decision seam."""
    _, db_path, _ = profile_router

    profile = RoutingProfile(allowed_model_ids={"nonexistent-model"})
    decision = route(_route_req(), profile="ghost", profile_obj=profile)

    assert decision.selected is None
    assert decision.candidates_considered == 0
    assert decision.profile == "ghost"

    # The allowlist resolves to an empty restrict_to set: nothing intersects.
    conn = sqlite3.connect(db_path)
    conn.row_factory = sqlite3.Row
    rows = dispatcher.load_candidates(conn, "coding_general")
    admin_deprecated = dispatcher._admin_excluded_models(conn)
    conn.close()
    restrict = dispatcher._restrict_to_from_profile(profile, rows)
    assert restrict == set()
    assert admin_deprecated == set()

    # Every catalog row carries a profile-era rejection reason. The flex twin
    # is caught one filter earlier (latency precedes the profile check in the
    # documented rejection_reason ordering); everything else lands on the
    # profile-specific reason.
    for row in rows:
        reason = rejection_reason(
            row, restrict_to=restrict, **_hard_filters()
        )
        if row["model_id"] == M1_FLEX:
            assert reason == "latency_class(flex)"
        else:
            assert reason == "profile_excluded"

    # And that reason, persisted through the production write path, lands on
    # the route_decisions row for the no-candidate decision.
    persist_route_decision(
        "route", classification=decision, rejected_reason="profile_excluded"
    )

    conn = sqlite3.connect(db_path)
    conn.row_factory = sqlite3.Row
    row = conn.execute(
        "SELECT rejected_reason, selected_model, profile FROM route_decisions "
        "ORDER BY id DESC LIMIT 1"
    ).fetchone()
    conn.close()
    assert row is not None
    assert row["selected_model"] is None
    assert row["profile"] == "ghost"
    assert row["rejected_reason"] is not None
    assert row["rejected_reason"].startswith("profile_excluded")


def test_debug_filter_lines_exist_for_profile_exclusions(profile_router, logbuf):
    """Every catalog row dropped by the allowlist gets a ``filter`` line.

    Known limitation, pinned as-is: route()'s debug loop calls
    ``rejection_reason(row, **filters)`` WITHOUT ``restrict_to``, so the
    logged reason for a profile-excluded row is the placeholder ``-`` rather
    than ``profile_excluded``. The authoritative reason (via
    ``rejection_reason(..., restrict_to=...)``) is pinned in the ghost
    test above; this test pins that the dropped rows at least surface in
    the debug log at all. A fix that threads restrict_to into the loop
    should update the ``== "-"`` expectations to ``== "profile_excluded"``.
    """
    _, _, _ = profile_router
    route(
        _route_req(),
        profile="ghost",
        profile_obj=RoutingProfile(allowed_model_ids={"nonexistent-model"}),
    )

    reasons: dict[str, str] = {}
    for line in logbuf.getvalue().splitlines():
        if line.startswith("filter "):
            fields = dict(
                token.split("=", 1) for token in line.split(" ")[1:] if "=" in token
            )
            reasons[fields.get("model", "")] = fields.get("reason", "")

    # Nothing survived, so every catalog row produced a filter line.
    for model_id in (M1, M1_FLEX, M2, LOCAL, LOCAL2):
        assert model_id in reasons, model_id
    # The flex row is dropped by the latency filter (reason visible), the
    # rest by the allowlist (reason currently '-').
    assert reasons[M1_FLEX] == "latency_class(flex)"
    for model_id in (M1, M2, LOCAL, LOCAL2):
        assert reasons[model_id] == "-"


def test_restriction_set_passes_rows_inside_the_allowlist(profile_router):
    """The same restrict_to set keeps rows it names — the set is precise, not
    an unconditional kill switch."""
    _, db_path, _ = profile_router

    profile = RoutingProfile(allowed_model_ids={M1, M2})
    conn = sqlite3.connect(db_path)
    conn.row_factory = sqlite3.Row
    rows = dispatcher.load_candidates(conn, "coding_general")
    conn.close()

    restrict = dispatcher._restrict_to_from_profile(profile, rows)
    assert restrict == {M1, M2}

    survivors = select_candidates(rows, restrict_to=restrict, **_hard_filters())
    assert {r["model_id"] for r in survivors} == {M1, M2}


# =============================================================================
# Plank (d): Profiles AND with circuit-breaker, admin overrides,
#            and eligible_categories.
# =============================================================================


def test_profile_excludes_circuit_open_model(profile_router, monkeypatch):
    """A profile that allows only m1 must still drop it when the circuit
    breaker has it open, and the rejection reason is the exclusion reason."""
    _, _, _ = profile_router

    monkeypatch.setattr(dispatcher.cfg.circuit_breaker, "enabled", True)
    circuit_breaker.record_failure(
        M1,
        "neuralwatt",
        initial_cooldown=dispatcher.cfg.circuit_breaker.initial_cooldown_seconds,
        max_cooldown=dispatcher.cfg.circuit_breaker.max_cooldown_seconds,
        backoff_multiplier=dispatcher.cfg.circuit_breaker.backoff_multiplier,
    )

    profile = RoutingProfile(allowed_model_ids={M1})
    decision = route(_route_req(), profile="m1-only", profile_obj=profile)
    assert decision.selected is None
    assert decision.candidates_considered == 0

    # The composition is visible at the primitive level too: the exclude_models
    # set (circuit-open rows) drops m1 even though the profile admits it.
    row = _model_row(M1)
    assert (
        rejection_reason(row, restrict_to={M1}, **_hard_filters(exclude_models={M1}))
        == "circuit_open"
    )


def test_profile_excludes_admin_deprecated_model(profile_router):
    """A profile that allows only m1 must drop it when an admin override
    deprecates it; the survivor set is empty."""
    _, db_path, _ = profile_router

    conn = sqlite3.connect(db_path)
    conn.execute(
        "INSERT INTO admin_model_overrides "
        "(model_id, provider, availability, reason, updated_at) "
        "VALUES (?, 'neuralwatt', 'deprecated', 'test', "
        "'2026-08-22T00:00:00+00:00')",
        (M1,),
    )
    conn.commit()
    conn.close()

    profile = RoutingProfile(allowed_model_ids={M1})
    decision = route(_route_req(), profile="m1-only", profile_obj=profile)

    assert decision.selected is None
    assert decision.candidates_considered == 0


def test_profile_excludes_category_ineligible_model(profile_router):
    """A profile that allows only m1 must drop it when m1's
    eligible_categories excludes the task's category."""
    _, db_path, _ = profile_router

    conn = sqlite3.connect(db_path)
    conn.execute(
        "UPDATE models SET eligible_categories = 'coding_refactor' "
        "WHERE model_id = ?",
        (M1,),
    )
    conn.commit()
    conn.close()

    profile = RoutingProfile(allowed_model_ids={M1})
    req = TaskRequest(
        task="summarize this",
        task_category="summarization",
        task_tier=2,
        required_context_tokens=100,
    )
    decision = route(req, profile="m1-only", profile_obj=profile)

    assert decision.selected is None
    assert decision.candidates_considered == 0

    # And the primitive composes: the category gate fires even though the
    # profile allowlist names the row.
    row = _model_row(M1, eligible_categories="coding_refactor")
    assert (
        rejection_reason(
            row, task_category="summarization", restrict_to={M1}, **_hard_filters()
        )
        == "category_ineligible"
    )


def test_profile_of_two_models_selects_the_survivor(profile_router):
    """When the profile allows both m1 and m2 and an admin override kills m1,
    the other candidate still wins — the filters compose without wiping the
    whole profile set."""
    _, db_path, _ = profile_router

    conn = sqlite3.connect(db_path)
    conn.execute(
        "INSERT INTO admin_model_overrides "
        "(model_id, provider, availability, reason, updated_at) "
        "VALUES (?, 'neuralwatt', 'deprecated', 'test', "
        "'2026-08-22T00:00:00+00:00')",
        (M1,),
    )
    conn.commit()
    conn.close()

    profile = RoutingProfile(allowed_model_ids={M1, M2})
    decision = route(_route_req(), profile="m1-or-m2", profile_obj=profile)

    assert decision.selected is not None
    assert decision.selected.model_id == M2
    assert decision.candidates_considered == 1


# =============================================================================
# Plank (e): Exploration picks only from profile-filtered candidates.
# =============================================================================


def _force_exploration(monkeypatch, *, seed: int = 0) -> None:
    monkeypatch.setattr(dispatcher.cfg.exploration, "enabled", True)
    monkeypatch.setattr(dispatcher.cfg.exploration, "epsilon", 1.0)
    monkeypatch.setattr(dispatcher.cfg.exploration, "max_tier", 3)
    # The local rows price above the cloud winner by more than the default
    # cap of 4.0; widen it so the explore path is reachable at all.
    monkeypatch.setattr(dispatcher.cfg.exploration, "max_cost_ratio", 1_000.0)
    monkeypatch.setattr(dispatcher, "_router_rng", random.Random(seed))


def test_exploration_never_leaves_the_locality_profile(profile_router, monkeypatch):
    """With exploration forced on, the locality profile restricts candidates
    to ollama-local rows, so every explored pick is local — exploration
    cannot reach a cloud row it never sees."""
    _, _, _ = profile_router
    _force_exploration(monkeypatch)

    profile = BUILTIN_PROFILES["locality"]
    seen: set[tuple[str, str]] = set()
    explored_any = False
    for _ in range(30):
        decision = route(_route_req(), profile="locality", profile_obj=profile)
        assert decision.selected is not None
        seen.add((decision.selected.model_id, decision.selected.provider))
        explored_any = explored_any or decision.explored

    assert seen, "the locality profile must route to something"
    assert explored_any, "with two local candidates and epsilon=1 the explore path runs"
    assert {model_id for model_id, _ in seen} <= {LOCAL, LOCAL2}
    assert all(provider == "ollama-local" for _, provider in seen)


def test_exploration_default_profile_can_reach_a_cloud_row(
    profile_router, monkeypatch
):
    """Positive control: same rng, same epsilon, default profile — exploration
    does pick a cloud row (the zero-sample one). This is what makes the
    locality assertion above meaningful rather than vacuous."""
    _, _, _ = profile_router
    _force_exploration(monkeypatch)

    seen: set[tuple[str, str]] = set()
    for _ in range(30):
        decision = route(
            _route_req(), profile="default", profile_obj=BUILTIN_PROFILES["default"]
        )
        assert decision.selected is not None
        seen.add((decision.selected.model_id, decision.selected.provider))

    assert any(provider == "neuralwatt" for _, provider in seen), seen
    assert any(model_id == M2 for model_id, _ in seen), (
        "the zero-sample cloud row is exactly what exploration targets"
    )


def test_exploration_stays_inside_an_allowed_model_ids_profile(
    profile_router, monkeypatch
):
    """An allowed_model_ids profile pins exploration to its allowlist."""
    _, _, _ = profile_router
    _force_exploration(monkeypatch)

    allowlist = {M1, M2}
    profile = RoutingProfile(allowed_model_ids=allowlist)
    for _ in range(30):
        decision = route(_route_req(), profile="m1-or-m2", profile_obj=profile)
        assert decision.selected is not None
        assert decision.selected.model_id in allowlist


# =============================================================================
# Route endpoint round-trip: profiles resolve and persist end to end.
# =============================================================================


def test_route_endpoint_batch_profile_round_trip(profile_router):
    """/route with profile=batch stores profile and batch latency on the row."""
    client, db_path, _ = profile_router

    resp = client.post(
        "/route", json={"task": "write me a function", "profile": "batch"}
    )
    assert resp.status_code == 200
    data = resp.json()
    assert data["profile"] == "batch"
    assert data["latency_tolerance"] == "batch"
    assert data["selected"]["model_id"] == M1_FLEX

    conn = sqlite3.connect(db_path)
    conn.row_factory = sqlite3.Row
    row = conn.execute(
        "SELECT profile, latency_tolerance, selected_model "
        "FROM route_decisions ORDER BY id DESC LIMIT 1"
    ).fetchone()
    conn.close()
    assert row is not None
    assert row["profile"] == "batch"
    assert row["latency_tolerance"] == "batch"
    assert row["selected_model"] == M1_FLEX


def test_bare_auto_resolves_to_configured_default_profile(profile_router, monkeypatch):
    """A bare ``auto`` model fills from ``routing.default_profile``, not a hard-coded "default".

    Monkeypatching ``cfg.routing.default_profile = "batch"`` causes a route
    whose profile kwarg is the default to behave exactly like ``auto:batch``.
    """
    _, _, _ = profile_router
    monkeypatch.setattr(dispatcher.cfg.routing, "default_profile", "batch")

    decision = route(_route_req())
    assert decision.latency_tolerance == "batch"
    assert decision.selected is not None
    assert decision.selected.model_id == M1_FLEX
    assert decision.selected.latency_class == "flex"


def test_the_route_endpoint_honours_the_configured_default_profile(
    profile_router, monkeypatch
):
    """/route is the documented no-spend probe of "what would the router
    pick". It resolved the literal built-in named "default" rather than
    routing.default_profile, so it answered a different question than the
    router answers -- and wrote the wrong name into route_decisions.

    Found live: with default_profile = "MiMo Test", /route logged
    profile="default" and picked deepseek-v4-flash, while the same task
    through ``auto`` logged "MiMo Test" and picked xiaomi/mimo-v2.5. Read
    back from the decisions history, that looks like a broken profile
    overlay rather than a broken probe.
    """
    client, _, _ = profile_router
    monkeypatch.setattr(dispatcher.cfg.routing, "default_profile", "batch")

    resp = client.post("/route", json={"task": "write me a function"})

    assert resp.status_code == 200
    body = resp.json()
    assert body["profile"] == "batch"
    assert body["latency_tolerance"] == "batch"
    assert body["selected"]["model_id"] == M1_FLEX


def test_an_explicit_profile_still_wins_over_the_configured_default(
    profile_router, monkeypatch
):
    """The narrowing stays narrow: a caller naming a profile is deliberately
    probing that one, including the built-in called "default"."""
    client, _, _ = profile_router
    monkeypatch.setattr(dispatcher.cfg.routing, "default_profile", "batch")

    resp = client.post(
        "/route", json={"task": "write me a function", "profile": "default"}
    )

    assert resp.status_code == 200
    assert resp.json()["profile"] == "default"