FabledScribe/tests/test_rule_hint_band.py

"""An act surfaces a SET of rules, and rank decides how loudly (#3851).

WHY THIS EXISTS

`RULEHINT_LIMIT` was 1. That was right while retrieval merely SUPPLEMENTED a
33-rule resident set — one salient rule beside everything already loaded. It
stops being right the moment milestone 394 removes residency, because then
this arm is the whole delivery, and `git push origin dev` is governed by
rules 1, 2, 9 and 140 at once, each of which alone permits the mistake the
others catch.

Two instruments, and they answer different questions:

  - `_rule_band` decides HOW MANY. A fixed k fills its slots whether or not
    anything deserves them; a band keeps only what scored close to the top,
    so one clearly-relevant rule still shows one.
  - `compact` decides HOW LOUD. Measured at #3851: a full line is ~143 tokens
    once the trigger is rendered, so five of them cost ~646 before every Bash
    call. Top-full-plus-references costs ~198.

WHAT THIS PINS

Structure, never wording — the lines are prose and will be rewritten:

  1. The band keeps the top hit and everything within `_RULEHINT_BAND`, and
     drops what falls outside. Falsified below from both sides: a hit just
     inside survives, a hit just outside does not.
  2. Rank decides volume — the first line carries the trigger, later lines do
     not, and every line names its rule's id so any of them can be pulled.
  3. The band reads SCORES ONLY. A top hit the session has already seen still
     anchors the band, and its score still sets the cutoff. This is the axis
     independence the renderer already keeps between `kind` and `seen`, and
     the regression it prevents is subtle: letting the ledger reorder the
     band would make "you were told this" change what counts as relevant.

The band width itself is deliberately NOT pinned. It is a tuning value with
a comment recording the measurement behind it, and a test asserting 0.05
would fail on every future retune while proving nothing about behaviour —
so the cases below express their scores as offsets from the constant.
"""
import pytest

from scribe.services.plugin_context import (
    _RULEHINT_BAND,
    _rule_band,
    _rule_hint_line,
)
from tests.helpers import fake_rule

_TRIGGER = "about to run git push with an earlier CI run still unread"


def _hit(score: float, rule_id: int):
    return (score, fake_rule(id=rule_id, title=f"rule {rule_id}",
                             when_to_apply=_TRIGGER))


def test_an_empty_result_stays_empty():
    """No hits is not a crash and not a phantom line."""
    assert _rule_band([]) == []


def test_the_band_keeps_a_hit_just_inside_it():
    """The whole point: a close second rule reaches the agent."""
    top = 0.75
    hits = [_hit(top, 1), _hit(top - _RULEHINT_BAND + 0.01, 2)]
    assert [r.id for _s, r in _rule_band(hits)] == [1, 2]


def test_the_band_drops_a_hit_just_outside_it():
    """And the band must actually BIND, or it is a fixed k wearing a hat."""
    top = 0.75
    hits = [_hit(top, 1), _hit(top - _RULEHINT_BAND - 0.01, 2)]
    assert [r.id for _s, r in _rule_band(hits)] == [1]


def test_one_clearly_better_rule_still_surfaces_alone():
    """The behaviour the old limit of 1 got right, which must not regress.

    A moment with a single relevant rule shows one line, because the corpus
    said so — not because a constant capped it.
    """
    hits = [_hit(0.80, 1), _hit(0.55, 2), _hit(0.54, 3)]
    assert [r.id for _s, r in _rule_band(hits)] == [1]


def test_a_flat_cluster_surfaces_together():
    """Measured shape of this corpus: adjacent rules sit ~0.02 apart.

    Four rules governing one act is the `git push` case the step exists for,
    and at the measured spacing they must arrive together rather than the
    ranker picking one of four near-ties.
    """
    hits = [_hit(0.757, 2), _hit(0.735, 7), _hit(0.726, 1), _hit(0.711, 9)]
    assert [r.id for _s, r in _rule_band(hits)] == [2, 7, 1, 9]


def test_the_band_is_computed_from_scores_not_from_the_ledger():
    """A seen top hit still anchors the band (#3750 x #3851).

    `_rule_band` never learns what the session has seen — dedup happens after
    it, in the arms. Pinned here because the tempting "reorder so a fresh rule
    leads" would silently change the cutoff, and the failure is invisible: the
    arm would still emit lines, just the wrong set.
    """
    hits = [_hit(0.80, 1), _hit(0.78, 2), _hit(0.60, 3)]
    kept = _rule_band(hits)
    # Independent of any `already` set, because it is not consulted.
    assert [r.id for _s, r in kept] == [1, 2]


@pytest.mark.parametrize("seen", [False, True])
def test_the_leading_line_carries_the_trigger(seen):
    """Rank 0 gets the full rendering, on either tail."""
    line = _rule_hint_line(
        fake_rule(id=4, title="dev is home", when_to_apply=_TRIGGER),
        where="to this Bash call", seen=seen, compact=False,
    )
    assert _TRIGGER in line
    assert "get_rule(4)" in line


@pytest.mark.parametrize("seen", [False, True])
def test_a_later_line_cites_its_rule_without_quoting_the_trigger(seen):
    """Rank > 0 is a reference: identity and pointer, no trigger.

    The trigger is the expensive half — 300-400 characters after #3855 — and
    the leading line has already demonstrated the shape. Both assertions
    matter: dropping the trigger is the saving, and keeping `get_rule(id)` is
    what makes the saving safe, because a cited rule the reader cannot pull is
    just noise.
    """
    line = _rule_hint_line(
        fake_rule(id=4, title="dev is home", when_to_apply=_TRIGGER),
        where="to this Bash call", seen=seen, compact=True,
    )
    assert _TRIGGER not in line
    assert "dev is home" in line
    assert "get_rule(4)" in line


def test_shortening_a_line_does_not_decide_what_it_says_about_holding():
    """A reference still tells a repeat from a first surfacing (#3750 x #3851).

    This is the regression the first cut of #3851 actually shipped: the
    compact branch dropped the tail along with the trigger, so a rule the
    session had already been told read exactly like one it had not. #3750's
    whole argument is that the two are different claims — a repeat is rendered
    precisely because the session may no longer HOLD what it was told — and
    the tail is the entire difference a reader can act on.

    `compact` and the ledger are independent axes. How much room a line gets is
    a fact about its rank; what the session holds is a fact about the ledger;
    and neither may be allowed to answer the other's question.

    THREE STATES SINCE #4100, so this checks three. The axis grew and the test
    grew with it — pinning only two would leave the compact branch free to
    collapse the new middle state into either neighbour, which is the same
    regression this was written for with one more place to hide.
    """
    rule = fake_rule(id=4, title="dev is home", when_to_apply=_TRIGGER)
    fresh = _rule_hint_line(rule, where="here", seen=False, compact=True)
    named = _rule_hint_line(rule, where="here", seen=True, compact=True)
    held = _rule_hint_line(rule, where="here", seen=True, held=True, compact=True)

    assert len({fresh, named, held}) == 3, (
        "the compact branch collapsed two holding states into one line"
    )
    assert "not in this session's loaded set" in fresh
    assert "not opened" in named
    assert "no longer hold it" in held
    # The claim that most needs to survive shortening: a line the session never
    # opened must not imply it did, however little room the line was given.
    assert "no longer hold it" not in named


def test_a_compact_line_is_materially_shorter_than_a_full_one():
    """The cost claim, asserted rather than left in a comment.

    Not a token count — that would pin the tokenizer. Half the characters is
    the property that makes widening the arm affordable, and it is what fails
    if a later edit puts the trigger back into the compact branch.

    Measured against a REALISTIC trigger, because that is where the saving
    lives: the rules this arm carries run 300-400 characters of trigger after
    #3855, and a toy one-line trigger would make this pass on a compact branch
    that had stopped saving anything.
    """
    long_trigger = (
        "Opening or merging a `dev`->`main` pull request, running "
        "`git push origin main`, `git tag`, or minting a release, image tag "
        "or other public artifact. Also whenever CI has just gone green and "
        "the next step feels like shipping it, and whenever an earlier merge "
        "this session reads like standing permission for the next one."
    )
    rule = fake_rule(id=4, title="dev is home", when_to_apply=long_trigger)
    full = _rule_hint_line(rule, where="here", seen=False, compact=False)
    compact = _rule_hint_line(rule, where="here", seen=False, compact=True)
    assert len(compact) * 2 < len(full)


def test_a_preference_keeps_its_noun_when_compact():
    """Force survives the shortening (#3849).

    `kind` and `compact` are independent axes. A preference rendered as a
    reference must still not read as a rule — the noun is the whole of the
    visual difference, so losing it in the compact branch would make every
    cited preference bind.
    """
    line = _rule_hint_line(
        fake_rule(id=5, title="pace debugging", kind="preference",
                  when_to_apply=_TRIGGER),
        where="here", seen=False, compact=True,
    )
    assert "preference" in line.lower()
    assert "standing rule" not in line.lower()