Skip to content

Compress tool outputs, logs, files, and RAG chunks before they reach the LLM. 20% fewer tokens for coding agents, 60-95% fewer tokens for JSON, same answers. Library, proxy, MCP server.

50.5 KB
Raw Blame
"""Tests for headroom.proxy.output_savings — the counterfactual estimator."""

from __future__ import annotations

import json

import pytest

from headroom.proxy.output_savings import (
    MEASURED_MIN_CLUSTERS,
    MEASURED_SUPERSEDE_MIN_CLUSTERS,
    BaselineModel,
    SavingsLedger,
    SavingsRecorder,
    assign_arm,
    conversation_key_from_body,
    conversation_label,
    echo_ratio,
    input_bucket,
    ledger_lock,
    model_family,
    stratum_key,
    stratum_label,
)

# ---------------------------------------------------------------------------
# stratification primitives
# ---------------------------------------------------------------------------


# A treatment observation only counts when the request was actually shaped,
# evidenced by the shaper's own verbosity label on the same channel.
SHAPED = "output_shaper:verbosity:L2"


class TestStratification:
    def test_input_buckets_monotone(self):
        assert input_bucket(0) == "xs"
        assert input_bucket(1_999) == "xs"
        assert input_bucket(2_000) == "s"
        assert input_bucket(8_000) == "m"
        assert input_bucket(32_000) == "l"
        assert input_bucket(200_000) == "xl"

    def test_model_family_collapses_point_releases(self):
        assert model_family("claude-opus-4-8") == "opus"
        assert model_family("claude-opus-4-7") == "opus"
        assert model_family("claude-sonnet-4-6") == "sonnet"
        assert model_family("gpt-4o") == "gpt"
        assert model_family("something-weird") == "other"

    def test_stratum_key_is_most_to_least_specific(self):
        key = stratum_key(
            turn_kind="new_user_ask", input_tokens=5000, model="claude-opus-4-8", has_tools=True
        )
        assert key == "opus|new_user_ask|s|tools"

    def test_stratum_key_distinguishes_tools(self):
        a = stratum_key(turn_kind="x", input_tokens=100, model="m", has_tools=True)
        b = stratum_key(turn_kind="x", input_tokens=100, model="m", has_tools=False)
        assert a != b


# ---------------------------------------------------------------------------
# holdout arm assignment
# ---------------------------------------------------------------------------


class TestArmAssignment:
    def test_zero_holdout_always_treatment(self):
        assert assign_arm("anything", 0.0) == "treatment"

    def test_full_holdout_always_control(self):
        assert assign_arm("anything", 1.0) == "control"

    def test_assignment_is_stable_for_same_key(self):
        assert assign_arm("conv-123", 0.5) == assign_arm("conv-123", 0.5)

    def test_roughly_matches_fraction(self):
        keys = [f"conv-{i}" for i in range(4000)]
        control = sum(1 for k in keys if assign_arm(k, 0.1) == "control")
        # 10% holdout over 4000 keys — allow generous slack for hash noise.
        assert 250 < control < 550

    def test_conversation_key_stable_across_turns(self):
        first = {
            "model": "claude-opus-4-8",
            "messages": [{"role": "user", "content": "build a cache"}],
        }
        later = {
            "model": "claude-opus-4-8",
            "messages": [
                {"role": "user", "content": "build a cache"},
                {"role": "assistant", "content": "ok"},
                {"role": "user", "content": [{"type": "tool_result", "content": "x"}]},
            ],
        }
        assert conversation_key_from_body(first) == conversation_key_from_body(later)

    def test_conversation_key_differs_by_first_message(self):
        a = {"model": "m", "messages": [{"role": "user", "content": "task A"}]}
        b = {"model": "m", "messages": [{"role": "user", "content": "task B"}]}
        assert conversation_key_from_body(a) != conversation_key_from_body(b)

    def test_conversation_key_uses_responses_stable_metadata(self):
        a = {
            "model": "gpt-5",
            "client_metadata": {"session_id": "session-1"},
            "input": "task A",
        }
        b = {
            "model": "gpt-5",
            "client_metadata": {"session_id": "session-2"},
            "input": "task A",
        }
        assert conversation_key_from_body(a) != conversation_key_from_body(b)

    def test_conversation_key_does_not_use_responses_delta_text(self):
        user_turn = {
            "model": "gpt-5",
            "instructions": "same session instructions",
            "input": "task A",
        }
        tool_turn = {
            "model": "gpt-5",
            "instructions": "same session instructions",
            "input": [
                {
                    "type": "function_call_output",
                    "call_id": "call_1",
                    "output": "ok",
                }
            ],
        }
        assert conversation_key_from_body(user_turn) == conversation_key_from_body(tool_turn)

    def test_conversation_key_unwraps_ws_response_create(self):
        http_body = {"model": "gpt-5", "input": "build a cache"}
        ws_body = {
            "type": "response.create",
            "response": {"model": "gpt-5", "input": "build a cache"},
        }
        assert conversation_key_from_body(http_body) == conversation_key_from_body(ws_body)

    def test_conversation_key_uses_responses_conversation_id(self):
        a = {
            "model": "gpt-5",
            "conversation": "conv_1",
            "input": [
                {
                    "type": "message",
                    "role": "user",
                    "content": [{"type": "input_text", "text": "task A"}],
                }
            ],
        }
        b = {
            "model": "gpt-5",
            "conversation": "conv_2",
            "input": [
                {
                    "type": "message",
                    "role": "user",
                    "content": [{"type": "input_text", "text": "task B"}],
                }
            ],
        }
        assert conversation_key_from_body(a) != conversation_key_from_body(b)


# ---------------------------------------------------------------------------
# baseline model
# ---------------------------------------------------------------------------


class TestBaselineModel:
    def test_observe_and_lookup_exact(self):
        m = BaselineModel()
        for v in (100, 200, 300):
            m.observe("opus|new_user_ask|s|tools", v)
        mean, var, n = m.lookup("opus|new_user_ask|s|tools")
        assert mean == 200.0
        assert n == 3
        assert var > 0

    def test_lookup_backs_off_to_prefix(self):
        m = BaselineModel()
        m.observe("opus|new_user_ask|s|tools", 500)
        # Query a sibling stratum (different tools flag) — backs off on prefix.
        mean, _, n = m.lookup("opus|new_user_ask|s|notools")
        assert mean == 500.0
        assert n == 1

    def test_an_unobserved_family_is_not_scored_against_the_global_mean(self):
        # The baseline is seeded once, from what the user ran before installing.
        # Every family adopted later lands here, and the all-requests mean is
        # not a control for any of them.
        m = BaselineModel()
        m.observe("opus|a|s|tools", 100)
        m.observe("sonnet|b|m|notools", 300)
        assert m.lookup("gpt|totally|xl|tools") == (0.0, 0.0, 0)

    def test_the_global_mean_remains_available_on_request(self):
        m = BaselineModel()
        m.observe("opus|a|s|tools", 100)
        m.observe("sonnet|b|m|notools", 300)
        mean, _, n = m.lookup("gpt|totally|xl|tools", fall_back_to_global=True)
        assert mean == 200.0  # global mean of 100 and 300
        assert n == 2

    def test_prefix_backoff_merges_every_neighbour_not_the_first_one_hashed(self):
        # Taking the first matching stratum in dict order made the answer
        # depend on insertion order, so a round-tripped ledger could score the
        # same request differently.
        m = BaselineModel()
        m.observe("opus|ask|l|tools", 1000)
        m.observe("opus|ask|l|notools", 100)
        mean, _, n = m.lookup("opus|ask|l|other")
        assert (mean, n) == (550.0, 2)

        reversed_order = BaselineModel()
        reversed_order.observe("opus|ask|l|notools", 100)
        reversed_order.observe("opus|ask|l|tools", 1000)
        assert reversed_order.lookup("opus|ask|l|other") == m.lookup("opus|ask|l|other")

    def test_roundtrip_serialization(self):
        m = BaselineModel()
        for v in (10, 20, 30):
            m.observe("k|a|s|tools", v)
        m2 = BaselineModel.from_dict(m.to_dict())
        assert m2.lookup("k|a|s|tools") == m.lookup("k|a|s|tools")
        assert m2.total_samples == 3

    def test_merge_is_equivalent_to_observing_both_streams(self):
        # Merging two baselines must equal observing every value against one
        # model — same totals per stratum and same global fallback.
        a = BaselineModel()
        for v in (100, 200):
            a.observe("opus|new_user_ask|s|tools", v)
        b = BaselineModel()
        b.observe("opus|new_user_ask|s|tools", 300)
        b.observe("sonnet|unknown|m|notools", 50)

        a.merge(b)

        mean, _, n = a.lookup("opus|new_user_ask|s|tools")
        assert n == 3
        assert mean == 200.0  # (100 + 200 + 300) / 3
        assert a.total_samples == 4  # 3 + 1 across both strata

        reference = BaselineModel()
        for v in (100, 200, 300):
            reference.observe("opus|new_user_ask|s|tools", v)
        reference.observe("sonnet|unknown|m|notools", 50)
        assert a.to_dict() == reference.to_dict()


# ---------------------------------------------------------------------------
# synthetic-control estimate
# ---------------------------------------------------------------------------


class TestEstimateFromBaseline:
    def _ledger_with_baseline(self, baseline_val: float, n: int = 50) -> SavingsLedger:
        ledger = SavingsLedger()
        for _ in range(n):
            ledger.baseline.observe("opus|new_user_ask|s|tools", baseline_val)
        return ledger

    def test_positive_savings_when_treatment_below_baseline(self):
        ledger = self._ledger_with_baseline(1000.0)
        for _ in range(20):
            ledger.record("treatment", "opus|new_user_ask|s|tools", 700)
        est = ledger.estimate_from_baseline()
        assert est.kind == "estimated"
        assert est.n_requests == 20
        # 20 requests * (1000 - 700) = 6000 tokens saved.
        assert abs(est.tokens_saved - 6000) < 1e-6
        assert abs(est.pct - 30.0) < 1e-6

    def test_signed_delta_not_clamped(self):
        # A treatment request LARGER than baseline must reduce the total, not
        # be clamped to zero (clamping would bias the estimate upward).
        ledger = self._ledger_with_baseline(1000.0)
        ledger.record("treatment", "opus|new_user_ask|s|tools", 700)
        ledger.record("treatment", "opus|new_user_ask|s|tools", 1400)
        est = ledger.estimate_from_baseline()
        # (1000-700) + (1000-1400) = 300 - 400 = -100
        assert abs(est.tokens_saved - (-100)) < 1e-6

    def test_zero_baseline_samples_yields_zero(self):
        ledger = SavingsLedger()
        ledger.record("treatment", "opus|x|s|tools", 500)
        est = ledger.estimate_from_baseline()
        # No baseline at all -> global is empty -> nothing contributes.
        assert est.n_requests == 0
        assert est.tokens_saved == 0.0

    def test_ci_band_brackets_point_estimate(self):
        ledger = SavingsLedger()
        for v in (900, 1000, 1100):
            for _ in range(20):
                ledger.baseline.observe("opus|new_user_ask|s|tools", v)
        for v in (600, 700, 800):
            for _ in range(20):
                ledger.record("treatment", "opus|new_user_ask|s|tools", v)
        est = ledger.estimate_from_baseline()
        assert est.ci_low_pct <= est.pct <= est.ci_high_pct
        assert est.ci_low_pct < est.ci_high_pct  # nonzero band given spread


# ---------------------------------------------------------------------------
# A/B measured estimate
# ---------------------------------------------------------------------------


class TestEstimateExcludesUnobservedStrata:
    def test_requests_without_baseline_evidence_are_left_out(self):
        ledger = SavingsLedger()
        for _ in range(10):
            ledger.baseline.observe("opus|ask|l|tools", 1000)
            ledger.record("treatment", "opus|ask|l|tools", 800)
        # A family the baseline never saw, with far shorter replies. Scoring it
        # against the global mean would credit ~950 saved tokens per request.
        for _ in range(40):
            ledger.record("treatment", "sonnet|new_user_ask|m|notools", 50)

        est = ledger.estimate_from_baseline()
        assert est.n_requests == 10, "only the observed stratum is scored"
        assert abs(est.tokens_saved - 2000) < 1e-6  # 10 * (1000 - 800)
        assert abs(est.pct - 20.0) < 1e-6

    def test_per_request_savings_are_zero_without_evidence(self):
        ledger = SavingsLedger()
        ledger.baseline.observe("opus|ask|l|tools", 1000)
        assert ledger.baseline.lookup("sonnet|new_user_ask|m|notools") == (0.0, 0.0, 0)


class TestEstimateFromHoldout:
    def test_none_without_control_data(self):
        ledger = SavingsLedger()
        ledger.record("treatment", "opus|x|s|tools", 500)
        assert ledger.estimate_from_holdout() is None

    def test_measured_difference_of_means(self):
        ledger = SavingsLedger()
        for i in range(30):
            ledger.record("control", "opus|new_user_ask|s|tools", 1000, f"c{i}")
            ledger.record("treatment", "opus|new_user_ask|s|tools", 750, f"t{i}")
        est = ledger.estimate_from_holdout()
        assert est is not None
        assert est.kind == "measured"
        # 30 * (1000 - 750) = 7500 saved; 25% of the 1000 baseline.
        assert abs(est.tokens_saved - 7500) < 1e-6
        assert abs(est.pct - 25.0) < 1e-6

    def test_only_strata_present_in_both_arms_contribute(self):
        ledger = SavingsLedger()
        for i in range(10):
            ledger.record("control", "opus|a|s|tools", 1000, f"c{i}")
            ledger.record("treatment", "opus|a|s|tools", 800, f"t{i}")
        # Treatment-only stratum must not contribute (no control to compare).
        ledger.record("treatment", "opus|b|m|notools", 50, "t99")
        est = ledger.estimate_from_holdout()
        assert est is not None
        assert est.n_requests == 10

    def test_best_estimate_prefers_measured(self):
        ledger = SavingsLedger()
        for i in range(MEASURED_SUPERSEDE_MIN_CLUSTERS):
            ledger.baseline.observe("opus|a|s|tools", 1000)
            ledger.record("control", "opus|a|s|tools", 1000, f"c{i}")
            ledger.record("treatment", "opus|a|s|tools", 900, f"t{i}")
        assert ledger.best_estimate().kind == "measured"

    def test_best_estimate_falls_back_to_estimated(self):
        ledger = SavingsLedger()
        for _ in range(10):
            ledger.baseline.observe("opus|a|s|tools", 1000)
            ledger.record("treatment", "opus|a|s|tools", 900)
        assert ledger.best_estimate().kind == "estimated"

    def test_a_wide_band_does_not_take_over_even_at_full_size(self):
        """The width gate still bites once the cluster gate is cleared."""
        ledger = SavingsLedger()
        for i in range(200):
            ledger.baseline.observe("opus|a|s|tools", 1000)
            ledger.record("treatment", "opus|a|s|tools", 800, f"t{i}")
        for i in range(30):
            ledger.record("control", "opus|a|s|tools", (200, 1000, 5000)[i % 3], f"c{i}")

        measured = ledger.estimate_from_holdout()
        assert measured is not None, "the arm is spread over enough conversations"
        assert (measured.ci_high_pct - measured.ci_low_pct) / 2 > 10, "and too noisy to believe"
        assert ledger.best_estimate().kind == "estimated"

    def test_a_holdout_covering_a_corner_of_traffic_does_not_take_over(self):
        ledger = SavingsLedger()
        # One stratum measured cleanly, but it is a fraction of the traffic.
        for i in range(30):
            ledger.baseline.observe("opus|a|s|tools", 1000)
            ledger.record("control", "opus|a|s|tools", 1000, f"c{i}")
            ledger.record("treatment", "opus|a|s|tools", 900, f"t{i}")
        for i in range(200):
            ledger.baseline.observe("opus|b|xl|tools", 4000)
            ledger.record("treatment", "opus|b|xl|tools", 3000, f"b{i}")
        assert ledger.best_estimate().kind == "estimated"

    def test_a_few_long_conversations_do_not_take_the_headline(self):
        """Five control conversations clear the cluster gate, and their
        per-request band looks tight, but they are five draws, not 1,000."""
        ledger = SavingsLedger()
        for i in range(200):
            ledger.baseline.observe("opus|a|s|tools", 1000)
            ledger.record("treatment", "opus|a|s|tools", 800, f"t{i}")
        for conv in range(5):
            for j in range(200):
                ledger.record("control", "opus|a|s|tools", 900 + 100 * conv + j % 3, f"c{conv}")

        measured = ledger.estimate_from_holdout()
        assert measured is not None, "a real arm by the cluster gate"
        assert (measured.ci_high_pct - measured.ci_low_pct) / 2 < 10, "and a tight-looking band"
        assert ledger.best_estimate().kind == "estimated"

    def test_unlabelled_legacy_traffic_does_not_block_the_measurement(self):
        """Requests recorded before conversations were tracked can never be
        measured, so they must not count against the measurement's coverage."""
        ledger = SavingsLedger()
        for _ in range(2000):
            ledger.baseline.observe("opus|a|s|tools", 1000)
            ledger.record("treatment", "opus|a|s|tools", 800)
        for i in range(MEASURED_SUPERSEDE_MIN_CLUSTERS):
            ledger.record("control", "opus|a|s|tools", 1000 + i % 3, f"c{i}")
            ledger.record("treatment", "opus|a|s|tools", 900 + i % 3, f"t{i}")

        assert ledger.estimate_from_baseline().n_requests == 2030
        assert ledger.best_estimate().kind == "measured"

    def test_a_young_holdout_still_reports_without_a_baseline(self):
        """Below the supersede floor there is nothing to displace, so the
        cluster-gated measurement is reported rather than nothing."""
        ledger = SavingsLedger()
        for i in range(MEASURED_MIN_CLUSTERS):
            ledger.record("control", "opus|a|s|tools", 1000, f"c{i}")
            ledger.record("treatment", "opus|a|s|tools", 800, f"t{i}")
        assert ledger.best_estimate().kind == "measured"

    def test_a_corner_holdout_does_not_set_the_headline_without_a_baseline(self):
        ledger = SavingsLedger()
        for i in range(MEASURED_MIN_CLUSTERS):
            ledger.record("control", "opus|a|s|tools", 1000, f"c{i}")
            ledger.record("treatment", "opus|a|s|tools", 100, f"t{i}")
        # The rest of the traffic has no control to compare against.
        for i in range(200):
            ledger.record("treatment", "opus|b|xl|tools", 3000, f"b{i}")
        assert ledger.estimate_from_holdout() is not None
        assert ledger.best_estimate().kind != "measured"

    def test_measured_wins_without_a_baseline_to_fall_back_on(self):
        """A holdout-only deployment (no ``learn --verbosity`` run) still reports."""
        ledger = SavingsLedger()
        for i in range(50):
            ledger.record("control", "opus|a|s|tools", 1000, f"c{i}")
            ledger.record("treatment", "opus|a|s|tools", 800, f"t{i}")
        assert ledger.estimate_from_baseline().n_requests == 0
        assert ledger.best_estimate().kind == "measured"


class TestHoldoutClusterGate:
    """A stratum needs distinct CONVERSATIONS in both arms, not requests.

    Assignment is conversation-stable, so one long agent session is one draw.
    Counting its requests as independent is what let four control requests
    decide a fleet machine's headline reduction.
    """

    @staticmethod
    def _fill(ledger, *, conversations, per_conversation, control_tokens=1000, treat_tokens=800):
        for i in range(conversations):
            for _ in range(per_conversation):
                ledger.record("control", "opus|a|s|tools", control_tokens, f"c{i}")
                ledger.record("treatment", "opus|a|s|tools", treat_tokens, f"t{i}")

    def test_one_conversation_per_arm_does_not_qualify(self):
        ledger = SavingsLedger()
        # 2,500 requests an arm, all from one session each side: the shape that
        # produced a -1.6% "measured" number on a real ledger.
        self._fill(ledger, conversations=1, per_conversation=2_500)
        assert ledger.estimate_from_holdout() is None

    def test_enough_conversations_qualifies(self):
        ledger = SavingsLedger()
        self._fill(ledger, conversations=MEASURED_MIN_CLUSTERS, per_conversation=2)
        est = ledger.estimate_from_holdout()
        assert est is not None
        assert est.kind == "measured"

    def test_thin_control_arm_does_not_ride_on_a_thick_treatment_one(self):
        ledger = SavingsLedger()
        for i in range(50):
            ledger.record("treatment", "opus|a|s|tools", 800, f"t{i}")
        for _ in range(400):
            ledger.record("control", "opus|a|s|tools", 1000, "one-session")
        assert ledger.estimate_from_holdout() is None

    def test_best_estimate_falls_back_when_the_holdout_is_one_conversation(self):
        ledger = SavingsLedger()
        for i in range(20):
            ledger.baseline.observe("opus|a|s|tools", 1000)
            ledger.record("treatment", "opus|a|s|tools", 900, f"t{i}")
            ledger.record("control", "opus|a|s|tools", 1000, "one-session")
        assert ledger.best_estimate().kind == "estimated"

    def test_a_ledger_written_before_conversations_were_tracked_does_not_qualify(self):
        # No cluster data at all: unverifiable, so it cannot clear the gate.
        ledger = SavingsLedger()
        for _ in range(100):
            ledger.record("control", "opus|a|s|tools", 1000)
            ledger.record("treatment", "opus|a|s|tools", 800)
        assert ledger.estimate_from_holdout() is None

    def test_cluster_tracking_saturates(self):
        ledger = SavingsLedger()
        for i in range(500):
            ledger.record("treatment", "opus|a|s|tools", 800, f"t{i}")
        # Bounded: the count is only ever compared against a threshold, so the
        # ledger does not grow a set entry per conversation forever.
        assert ledger.treatment["opus|a|s|tools"].n_clusters <= 32
        assert ledger.treatment["opus|a|s|tools"].n_clusters >= MEASURED_MIN_CLUSTERS

    def test_conversation_survives_a_save_load_cycle(self, tmp_path):
        ledger = SavingsLedger()
        for i in range(MEASURED_MIN_CLUSTERS):
            ledger.record("control", "opus|a|s|tools", 1000, f"c{i}")
            ledger.record("treatment", "opus|a|s|tools", 800, f"t{i}")
        path = tmp_path / "savings.json"
        ledger.save(path)
        assert SavingsLedger.load(path).estimate_from_holdout() is not None

    def test_recorder_reads_the_conversation_off_the_label_channel(self, tmp_path):
        recorder = SavingsRecorder(tmp_path / "savings.json", flush_every=1)
        for i in range(MEASURED_MIN_CLUSTERS):
            key = conversation_key_from_body({"messages": [{"role": "user", "content": f"q{i}"}]})
            assert recorder.record_from_labels(
                [
                    "router:noop",
                    "output_shaper:verbosity:concise",
                    stratum_label("treatment", "opus|a|s|tools"),
                    conversation_label(key),
                ],
                800,
            )
            assert recorder.record_from_labels(
                [conversation_label(key + "control"), stratum_label("control", "opus|a|s|tools")],
                1000,
            )
        assert SavingsLedger.load(tmp_path / "savings.json").estimate_from_holdout() is not None

    def test_a_request_without_a_conversation_label_still_records(self, tmp_path):
        recorder = SavingsRecorder(tmp_path / "savings.json", flush_every=1)
        assert recorder.record_from_labels(
            [stratum_label("treatment", "opus|a|s|tools"), "output_shaper:verbosity:concise"], 800
        )
        ledger = SavingsLedger.load(tmp_path / "savings.json")
        assert ledger.treatment["opus|a|s|tools"].n == 1
        assert ledger.treatment["opus|a|s|tools"].n_clusters == 0

    # -- provenance: clusters vouch for labelled observations, nothing else ---

    @staticmethod
    def _legacy_ledger_dict(requests=2_500, control_tokens=1000, treat_tokens=2000):
        """An arm as an upgraded ledger holds it: totals, no conversations.

        Those requests could all be one conversation -- the exact case the
        cluster gate exists to exclude -- and nothing on disk can say.
        """
        return {
            # Shaped-only arms can predate conversation provenance.
            "shaped_only": True,
            "baseline": {"strata": {}},
            "treatment": {
                "opus|a|s|tools": {
                    "n": requests,
                    "sum": float(requests * treat_tokens),
                    "sumsq": float(requests * treat_tokens**2),
                }
            },
            "control": {
                "opus|a|s|tools": {
                    "n": requests,
                    "sum": float(requests * control_tokens),
                    "sumsq": float(requests * control_tokens**2),
                }
            },
        }

    def test_upgraded_legacy_traffic_never_joins_the_measured_arm(self, tmp_path):
        """Five fresh conversations qualify the STRATUM, not the back catalogue.

        Before this split the reload kept n/sum/sumsq and the new labelled
        observations only added clusters to the same accumulator, so the moment
        the gate opened all 2,500 unattributable requests an arm were measured
        too -- reporting -99.8% over 2,505 requests while the conversations
        actually observed showed no difference at all.
        """
        path = tmp_path / "savings.json"
        path.write_text(json.dumps(self._legacy_ledger_dict()))
        ledger = SavingsLedger.load(path)
        assert ledger.estimate_from_holdout() is None, "legacy traffic alone cannot qualify"

        for i in range(MEASURED_MIN_CLUSTERS):
            ledger.record("control", "opus|a|s|tools", 1000, f"c{i}")
            ledger.record("treatment", "opus|a|s|tools", 1000, f"t{i}")

        est = ledger.estimate_from_holdout()
        assert est is not None, "the labelled conversations are a real sample"
        # Only the labelled requests are measured, and they show no difference.
        assert est.n_requests == MEASURED_MIN_CLUSTERS
        assert est.tokens_saved == pytest.approx(0.0)
        assert est.pct == pytest.approx(0.0)
        # The totals survive for the estimated / modelled tiers and reporting.
        assert ledger.treatment["opus|a|s|tools"].n == 2_500 + MEASURED_MIN_CLUSTERS

    def test_the_qualified_subset_survives_a_save_load_cycle(self, tmp_path):
        """The split has to persist, or the next restart re-merges the arms."""
        path = tmp_path / "savings.json"
        path.write_text(json.dumps(self._legacy_ledger_dict()))
        ledger = SavingsLedger.load(path)
        for i in range(MEASURED_MIN_CLUSTERS):
            ledger.record("control", "opus|a|s|tools", 1000, f"c{i}")
            ledger.record("treatment", "opus|a|s|tools", 1000, f"t{i}")
        ledger.save(path)

        reloaded = SavingsLedger.load(path)
        est = reloaded.estimate_from_holdout()
        assert est is not None
        assert est.n_requests == MEASURED_MIN_CLUSTERS
        assert est.tokens_saved == pytest.approx(0.0)
        assert reloaded.treatment["opus|a|s|tools"].n == 2_500 + MEASURED_MIN_CLUSTERS

    def test_later_unlabelled_requests_stay_out_of_a_qualified_stratum(self):
        """Qualifying a stratum does not open it to unattributable traffic."""
        ledger = SavingsLedger()
        for i in range(MEASURED_MIN_CLUSTERS):
            ledger.record("control", "opus|a|s|tools", 1000, f"c{i}")
            ledger.record("treatment", "opus|a|s|tools", 1000, f"t{i}")
        before = ledger.estimate_from_holdout()
        assert before is not None

        for _ in range(2_000):
            ledger.record("treatment", "opus|a|s|tools", 5)

        after = ledger.estimate_from_holdout()
        assert after is not None
        assert after.n_requests == before.n_requests
        assert after.tokens_saved == pytest.approx(before.tokens_saved)


# ---------------------------------------------------------------------------
# persistence
# ---------------------------------------------------------------------------


class TestLedgerPersistence:
    def test_roundtrip(self, tmp_path):
        ledger = SavingsLedger()
        ledger.baseline.observe("opus|a|s|tools", 1000)
        for i in range(MEASURED_MIN_CLUSTERS):
            ledger.record("treatment", "opus|a|s|tools", 800, f"t{i}")
            ledger.record("control", "opus|a|s|tools", 1000, f"c{i}")
        path = tmp_path / "savings.json"
        ledger.save(path)
        loaded = SavingsLedger.load(path)
        assert loaded.estimate_from_baseline().tokens_saved == (
            ledger.estimate_from_baseline().tokens_saved
        )
        assert loaded.estimate_from_holdout() is not None

    def test_load_missing_returns_empty(self, tmp_path):
        ledger = SavingsLedger.load(tmp_path / "nope.json")
        assert ledger.baseline.total_samples == 0

    def test_load_corrupt_returns_empty(self, tmp_path):
        p = tmp_path / "bad.json"
        p.write_text("{not json")
        ledger = SavingsLedger.load(p)
        assert ledger.baseline.total_samples == 0

    @staticmethod
    def _pre_rule_ledger(path):
        """A ledger written before shaped-only recording: arms, no marker."""
        baseline = BaselineModel()
        baseline.observe("opus|a|s|tools", 1000)
        legacy = {
            "baseline": baseline.to_dict(),
            "treatment": {"opus|a|s|tools": {"n": 3, "sum": 2400.0, "sumsq": 1920000.0}},
            "control": {},
        }
        path.write_text(json.dumps(legacy))
        return legacy

    @staticmethod
    def _legacy_warnings(caplog):
        return [r for r in caplog.records if "predates shaped-only" in r.getMessage()]

    def test_pre_rule_ledger_warns_once_and_is_rewritten(self, tmp_path, caplog):
        """/stats reloads the ledger on every poll. With the shaper disabled
        nothing ever flushes, so the legacy file used to survive and warn on
        every poll forever."""
        import logging

        path = tmp_path / "output_savings.json"
        legacy = self._pre_rule_ledger(path)

        with caplog.at_level(logging.WARNING):
            recorder = SavingsRecorder(str(path))
            for _ in range(3):
                recorder.estimate()

        assert len(self._legacy_warnings(caplog)) == 1
        on_disk = json.loads(path.read_text())
        assert on_disk["shaped_only"] is True
        assert on_disk["treatment"] == {}
        assert on_disk["baseline"] == legacy["baseline"]

    def test_stateless_pre_rule_ledger_warns_once_and_is_left_alone(self, tmp_path, caplog):
        import logging

        from headroom import paths

        path = tmp_path / "output_savings.json"
        legacy = self._pre_rule_ledger(path)

        paths.set_process_stateless(True)
        try:
            with caplog.at_level(logging.WARNING):
                recorder = SavingsRecorder(str(path))
                for _ in range(3):
                    recorder.estimate()
        finally:
            paths.set_process_stateless(False)

        assert len(self._legacy_warnings(caplog)) == 1
        assert json.loads(path.read_text()) == legacy
        assert not (tmp_path / "output_savings.json.lock").exists()

    def test_migration_keeps_a_baseline_learned_after_the_first_load(self, tmp_path, monkeypatch):
        """learn replaced the file between the recorder's first load and its
        migration write; the stale legacy baseline must not overwrite it."""
        path = tmp_path / "output_savings.json"
        self._pre_rule_ledger(path)
        learned = BaselineModel()
        for output_tokens in (400, 420, 380):
            learned.observe("opus|b|s|tools", output_tokens)

        real_load = SavingsLedger.load.__func__
        loads = []

        def load_then_learn(cls, p, **kwargs):
            ledger = real_load(cls, p, **kwargs)
            loads.append(p)
            if len(loads) == 1:
                SavingsLedger.store_baseline(path, learned)
            return ledger

        monkeypatch.setattr(SavingsLedger, "load", classmethod(load_then_learn))
        recorder = SavingsRecorder(str(path))

        on_disk = json.loads(path.read_text())
        assert on_disk["baseline"] == learned.to_dict()
        assert recorder._ledger.baseline.to_dict() == learned.to_dict()

    def test_migration_holds_the_ledger_lock_while_writing(self, tmp_path, monkeypatch):
        import threading

        path = tmp_path / "output_savings.json"
        self._pre_rule_ledger(path)
        real_save = SavingsLedger.save
        contended = []

        def probing_save(self, p):
            acquired = threading.Event()

            def other_writer():
                with ledger_lock(p):
                    acquired.set()

            probe = threading.Thread(target=other_writer)
            probe.start()
            contended.append(not acquired.wait(timeout=0.3))
            real_save(self, p)
            probe_threads.append(probe)

        probe_threads: list[threading.Thread] = []
        monkeypatch.setattr(SavingsLedger, "save", probing_save)
        SavingsRecorder(str(path))
        for probe in probe_threads:
            probe.join(timeout=5)

        assert contended == [True]
        assert json.loads(path.read_text())["shaped_only"] is True

    def test_migration_keeps_a_baseline_learned_during_the_locked_reload(
        self, tmp_path, monkeypatch
    ):
        """learn writes while the migration holds its re-read snapshot; it
        must wait for the lock rather than be overwritten by that snapshot."""
        import threading

        path = tmp_path / "output_savings.json"
        legacy = self._pre_rule_ledger(path)
        learned = BaselineModel()
        for output_tokens in (400, 420, 380):
            learned.observe("opus|b|s|tools", output_tokens)

        real_load = SavingsLedger.load.__func__
        loads = []
        writers: list[threading.Thread] = []

        def learn_during_second_load(cls, p, **kwargs):
            ledger = real_load(cls, p, **kwargs)
            loads.append(p)
            if len(loads) == 2:
                writer = threading.Thread(target=SavingsLedger.store_baseline, args=(path, learned))
                writer.start()
                writer.join(timeout=0.3)
                writers.append(writer)
            return ledger

        monkeypatch.setattr(SavingsLedger, "load", classmethod(learn_during_second_load))
        recorder = SavingsRecorder(str(path))
        assert len(writers) == 1
        writers[0].join(timeout=5)
        assert not writers[0].is_alive()

        on_disk = json.loads(path.read_text())
        assert on_disk["shaped_only"] is True
        assert on_disk["baseline"] == learned.to_dict()
        assert on_disk["baseline"] != legacy["baseline"]
        recorder.estimate()
        assert recorder._ledger.baseline.to_dict() == learned.to_dict()

    def test_store_baseline_waits_for_the_ledger_lock(self, tmp_path):
        """learn must not replace the file while another writer holds the lock
        between its re-read and its save."""
        import threading

        path = tmp_path / "output_savings.json"
        legacy = self._pre_rule_ledger(path)
        learned = BaselineModel()
        learned.observe("opus|b|s|tools", 400)

        writer = threading.Thread(target=SavingsLedger.store_baseline, args=(path, learned))
        with ledger_lock(path):
            writer.start()
            writer.join(timeout=0.3)
            assert writer.is_alive()
            assert json.loads(path.read_text()) == legacy
        writer.join(timeout=5)

        assert not writer.is_alive()
        assert json.loads(path.read_text())["baseline"] == learned.to_dict()


# ---------------------------------------------------------------------------
# echo ratio (direct waste signal)
# ---------------------------------------------------------------------------


class TestEchoRatio:
    def test_full_echo(self):
        ctx = "the quick brown fox jumps over the lazy dog every single time"
        assert echo_ratio(ctx, ctx, n=4) == 1.0

    def test_no_echo(self):
        out = "completely unrelated words appearing nowhere within the given source context here"
        ctx = "alpha beta gamma delta epsilon zeta eta theta iota kappa lambda"
        assert echo_ratio(out, ctx, n=4) == 0.0

    def test_partial_echo_between_zero_and_one(self):
        ctx = "alpha beta gamma delta epsilon zeta eta theta"
        out = "alpha beta gamma delta brand new tokens here now"
        r = echo_ratio(out, ctx, n=4)
        assert 0.0 < r < 1.0

    def test_short_output_returns_zero(self):
        assert echo_ratio("a b", "a b c d e f g h", n=8) == 0.0


# ---------------------------------------------------------------------------
# recorder baseline reload (learn-while-running)
# ---------------------------------------------------------------------------


class TestRecorderBaselineReload:
    """The recorder must pick up a baseline that ``learn --verbosity --apply``
    writes while the proxy is already running, and a flush must never overwrite
    that learned baseline with the recorder's own empty in-memory copy."""

    @staticmethod
    def _key() -> str:
        return SAMPLE_KEY

    def test_adopts_baseline_learned_after_start(self, tmp_path):
        path = str(tmp_path / "output_savings.json")
        key = self._key()

        recorder = SavingsRecorder(path, flush_every=1)
        for output_tokens in (200, 210, 190):
            recorder.record_from_labels([stratum_label("treatment", key), SHAPED], output_tokens)

        # No baseline to compare against yet, so there is nothing to estimate.
        assert recorder.estimate().n_requests == 0

        # Simulate `learn --verbosity --apply` writing a baseline to the same
        # file while the recorder is live (no restart).
        learned = SavingsLedger.load(path)
        for output_tokens in (400, 420, 380, 410):
            learned.baseline.observe(key, output_tokens)
        learned.save(path)

        estimate = recorder.estimate()
        assert estimate.n_requests > 0
        assert estimate.kind == "estimated"
        assert estimate.tokens_saved > 0

    def test_flush_does_not_clobber_learned_baseline(self, tmp_path):
        path = str(tmp_path / "output_savings.json")
        key = self._key()

        # Recorder starts before any baseline exists, so its in-memory baseline
        # is empty.
        recorder = SavingsRecorder(path, flush_every=1)

        learned = SavingsLedger.load(path)
        for output_tokens in (400, 420, 380, 410):
            learned.baseline.observe(key, output_tokens)
        learned.save(path)
        assert SavingsLedger.load(path).baseline.total_samples == 4

        recorder.record_from_labels([stratum_label("treatment", key), SHAPED], 200)
        recorder.flush()

        # The flush must keep the learned baseline rather than writing the empty
        # in-memory one over it.
        assert SavingsLedger.load(path).baseline.total_samples == 4

    def test_store_baseline_keeps_live_arms(self, tmp_path):
        path = str(tmp_path / "output_savings.json")
        key = self._key()

        recorder = SavingsRecorder(path, flush_every=1)
        recorder.record_from_labels([stratum_label("treatment", key), SHAPED], 200)

        learned = BaselineModel()
        learned.observe(key, 400)
        SavingsLedger.store_baseline(path, learned)

        on_disk = SavingsLedger.load(path)
        assert on_disk.baseline.to_dict() == learned.to_dict()
        assert on_disk.treatment[key].n == 1

    def test_does_not_downgrade_to_empty_disk_baseline(self, tmp_path):
        path = str(tmp_path / "output_savings.json")
        key = self._key()

        # Recorder already holds a learned baseline in memory.
        recorder = SavingsRecorder(path, flush_every=1)
        recorder._ledger.baseline.observe(key, 400)
        recorder._ledger.baseline.observe(key, 420)
        assert recorder._ledger.baseline.total_samples == 2

        # A stale/empty file on disk must not erase a baseline we already hold.
        SavingsLedger().save(path)
        recorder.flush()

        assert recorder._ledger.baseline.total_samples == 2

    def test_relearn_with_same_sample_count_is_adopted(self, tmp_path):
        path = str(tmp_path / "output_savings.json")
        key = self._key()

        recorder = SavingsRecorder(path, flush_every=1)
        for output_tokens in (200, 210, 190):
            recorder.record_from_labels([stratum_label("treatment", key), SHAPED], output_tokens)

        # First learn writes a baseline; the recorder adopts it.
        first = SavingsLedger.load(path)
        for output_tokens in (400, 400, 400, 400):
            first.baseline.observe(key, output_tokens)
        first.save(path)
        baseline_tokens_v1 = recorder.estimate().baseline_tokens
        assert baseline_tokens_v1 > 0

        # Re-running learn replaces the baseline in place with the SAME number of
        # samples but different values. A sample-count guard would miss this; the
        # recorder must still pick the new baseline up.
        relearned = SavingsLedger.load(path)
        relearned.baseline = BaselineModel()
        for output_tokens in (800, 800, 800, 800):
            relearned.baseline.observe(key, output_tokens)
        relearned.save(path)

        assert recorder.estimate().baseline_tokens > baseline_tokens_v1


# ---------------------------------------------------------------------------
# flush durability + event-loop safety
# ---------------------------------------------------------------------------

# Deterministic stratum key shared by the recorder tests below.
SAMPLE_KEY = stratum_key(
    turn_kind="code",
    input_tokens=8000,
    model="claude-opus-4-8",
    has_tools=True,
)


class TestFlushDurability:
    def test_crash_mid_write_leaves_previous_ledger_intact(self, tmp_path, monkeypatch):
        import headroom.fsutil

        path = str(tmp_path / "output_savings.json")
        key = SAMPLE_KEY

        recorder = SavingsRecorder(path, flush_every=1)
        recorder.record_from_labels([stratum_label("treatment", key), SHAPED], 200)
        recorder.flush()
        assert SavingsLedger.load(path).treatment[key].n == 1

        def _die_before_rename(*args, **kwargs):
            raise OSError(5, "simulated crash before rename")

        monkeypatch.setattr(headroom.fsutil.os, "replace", _die_before_rename)
        recorder.record_from_labels([stratum_label("treatment", key), SHAPED], 210)
        recorder.flush()  # OSError swallowed by the recorder — fail-open by design

        # The pre-crash sample must survive and no temp residue may be left
        # behind: a failed save may not corrupt or clutter the ledger.
        assert SavingsLedger.load(path).treatment[key].n == 1
        assert not list(tmp_path.glob("*.tmp"))

    def test_corrupt_ledger_warns_and_starts_empty(self, tmp_path, caplog):
        import logging

        path = tmp_path / "output_savings.json"
        path.write_text("{not json")

        with caplog.at_level(logging.WARNING):
            SavingsRecorder(str(path))

        assert caplog.records, "corrupt ledger was swallowed silently"

    def test_emit_request_outcome_flushes_off_the_loop_thread(self, tmp_path, monkeypatch):
        import asyncio
        import threading

        from headroom.proxy.outcome import RequestOutcome, emit_request_outcome

        path = str(tmp_path / "output_savings.json")
        recorder = SavingsRecorder(path, flush_every=1)
        monkeypatch.setattr("headroom.proxy.output_savings.get_recorder", lambda: recorder)

        saved_on_threads = []
        real_save = SavingsLedger.save

        def _spy_save(self, save_path):
            saved_on_threads.append(threading.get_ident())
            real_save(self, save_path)

        monkeypatch.setattr(SavingsLedger, "save", _spy_save)

        class _Metrics:
            async def record_request(self, **kwargs):
                pass

        class _Handler:
            def __init__(self):
                self.metrics = _Metrics()
                self.cost_tracker = None
                self.logger = None

        outcome = RequestOutcome(
            request_id="req-shaper",
            provider="openai",
            model="gpt-5",
            status_code=200,
            original_tokens=100,
            optimized_tokens=80,
            output_tokens=50,
            tokens_saved=20,
            attempted_input_tokens=100,
            transforms_applied=(stratum_label("treatment", SAMPLE_KEY), SHAPED),
        )
        asyncio.run(emit_request_outcome(_Handler(), outcome))

        loop_thread = threading.get_ident()
        assert saved_on_threads, "flush never ran"
        assert all(t != loop_thread for t in saved_on_threads)


class TestModelledTier:
    """The fallback for a deployment with no counterfactual of its own.

    The factor table ships EMPTY: open-source Headroom applies steering but
    does not claim a savings figure it has not measured. Factors arrive either
    from a holdout (which outranks this tier entirely) or from an extension
    calling ``register_modelled_factors``. These tests therefore register their
    own factors and restore the table afterwards -- they exercise the
    arithmetic, which is permanent, not the numbers, which are not.
    """

    @staticmethod
    @pytest.fixture
    def factors():
        """Install factors for level 3, then restore the real table."""
        from headroom.proxy.output_savings import (
            MODELLED_REDUCTION,
            register_modelled_factors,
        )

        saved = dict(MODELLED_REDUCTION)
        register_modelled_factors(3, 0.20, 0.40)
        try:
            yield (0.20, 0.40)
        finally:
            MODELLED_REDUCTION.clear()
            MODELLED_REDUCTION.update(saved)

    @staticmethod
    def _ledger_with(observed_total: int, n: int):
        from headroom.proxy.output_savings import SavingsLedger, stratum_key

        ledger = SavingsLedger()
        key = stratum_key(
            turn_kind="new_user_ask", input_tokens=1000, model="claude-sonnet-5", has_tools=False
        )
        for _ in range(n):
            ledger.record("treatment", key, observed_total // n)
        return ledger

    def test_ships_empty_so_an_unmeasured_deployment_claims_nothing(self):
        """No factors by default -> no modelled estimate, at any level.

        The dash this produces is the point: it is the correct rendering of
        "not measured". A built-in constant would be a number nobody measured
        on this deployment's traffic, which is the failure mode the tiering
        exists to prevent.
        """
        from headroom.proxy.output_savings import MODELLED_REDUCTION

        assert MODELLED_REDUCTION == {}
        led = self._ledger_with(5_000, 5)
        assert all(led.estimate_from_model(lv) is None for lv in (1, 2, 3, 4))

    def test_registering_factors_enables_the_tier(self, factors):
        assert self._ledger_with(5_000, 5).estimate_from_model(3) is not None

    def test_nonsense_factors_are_rejected_at_registration(self):
        """r=0 and r=1 break the r/(1-r) inversion; catch it at the door."""
        from headroom.proxy.output_savings import register_modelled_factors

        for bad in ((0.0, 0.4), (1.0, 1.0), (-0.1, 0.4), (0.5, 1.2)):
            with pytest.raises(ValueError):
                register_modelled_factors(3, *bad)
        with pytest.raises(ValueError, match="exceeds optimistic"):
            register_modelled_factors(3, 0.5, 0.2)

    def test_saving_inverts_the_reduction_rather_than_scaling_by_it(self, factors):
        """Observed output is POST-shaping, so saved is observed*r/(1-r).

        The naive observed*r understates the saving. This is the single
        arithmetic mistake the tier can make, so it is pinned.

        r is read from the table rather than hardcoded: the factors are
        re-measured whenever the steering text changes, and a test that
        snapshots them fails on every remeasure while testing nothing about
        the arithmetic it exists to protect.
        """
        from headroom.proxy.output_savings import MODELLED_REDUCTION

        ledger = self._ledger_with(10_000, 10)
        est = ledger.estimate_from_model(3)
        assert est is not None
        r = MODELLED_REDUCTION[3][0]
        assert 0 < r < 1, "a reduction factor outside (0,1) makes the inversion nonsense"
        assert est.tokens_saved == pytest.approx(10_000 * r / (1 - r), rel=1e-6)
        assert est.tokens_saved > 10_000 * r, "naive scaling would understate"
        # baseline = what the unshaped run would have emitted
        assert est.baseline_tokens == pytest.approx(10_000 + est.tokens_saved, rel=1e-6)

    def test_kind_is_modelled_so_the_ui_can_refuse_to_call_it_a_ci(self, factors):
        est = self._ledger_with(5_000, 5).estimate_from_model(3)
        assert est is not None and est.kind == "modelled"

    def test_band_is_the_two_provider_spread(self, factors):
        from headroom.proxy.output_savings import MODELLED_REDUCTION

        low, high = MODELLED_REDUCTION[3]
        est = self._ledger_with(5_000, 5).estimate_from_model(3)
        assert est is not None
        assert est.ci_low_pct == pytest.approx(low * 100)
        assert est.ci_high_pct == pytest.approx(high * 100)
        assert low <= high, "conservative end must not exceed the optimistic one"
        assert est.pct == est.ci_low_pct, "headline uses the conservative end"

    def test_unbenchmarked_level_yields_nothing_rather_than_a_guess(self):
        assert self._ledger_with(5_000, 5).estimate_from_model(1) is None

    def test_no_traffic_yields_nothing(self):
        from headroom.proxy.output_savings import SavingsLedger

        assert SavingsLedger().estimate_from_model(3) is None

    def test_a_real_baseline_supersedes_the_model(self):
        """The modelled tier is last resort; a learned baseline outranks it."""
        from headroom.proxy.output_savings import BaselineModel, SavingsLedger, stratum_key

        key = stratum_key(
            turn_kind="new_user_ask", input_tokens=1000, model="claude-sonnet-5", has_tools=False
        )
        baseline = BaselineModel()
        for _ in range(50):
            baseline.observe(key, 2000)
        ledger = SavingsLedger(baseline=baseline)
        for _ in range(10):
            ledger.record("treatment", key, 1000)
        assert ledger.best_estimate(3).kind == "estimated"

    def test_without_a_level_behaviour_is_unchanged(self):
        """Existing callers that pass no level must not silently gain a number."""
        est = self._ledger_with(5_000, 5).best_estimate()
        assert est.kind == "estimated" and est.n_requests == 0