core/generate/exhaustion.py
Shay 9d31f80fc8
fix(W-011/W-012): propagate recognition refusal + catch InnerLoopExhaustion (#258)
W-011: recognition refusal_reason now materializes in
CognitiveTurnResult.refusal_reason via RECOGNITION_REFUSED enum value.
Precedence: recognition wins over generation (earlier-fail boundary).

W-012: ChatRuntime.chat() catches InnerLoopExhaustion from generate()
and returns a typed refusal ChatResponse with refusal_reason populated,
instead of propagating as an unhandled exception.

Adds RefusalReason.RECOGNITION_REFUSED to generate/exhaustion.py.

Lane SHAs: 7/7 match (demos don't exercise refusal paths — no re-pin).
Smoke + cognition suites green. Full suite not run to completion.
2026-05-24 20:46:46 -07:00

134 lines
5.6 KiB
Python

"""
InnerLoopExhaustion — evidence-carrying honest refusal at the walk site.
ADR-0024 Phase 2. When the inner-loop admissibility check leaves no
admissible destination for the next step, the walk must refuse rather
than silently relax the constraint. Phase 1 raised a plain
``ValueError`` with a message string; Phase 2 promotes that to a
typed exception that carries the refusal evidence (region label,
step index, rejected attempts, machine-readable reason) so downstream
layers can materialise the refusal into trace evidence without
re-parsing the message.
This is *not* normalization or repair (CLAUDE.md §Normalization Rules).
The walk still refuses at the same site, with the same algebra, with
the same effect on caller control flow. The exception subclasses
``ValueError`` so every existing ``except ValueError`` handler in the
runtime/eval code paths continues to work byte-for-byte — the carrying
of structured evidence is additive.
Reason codes:
``INNER_LOOP_EXHAUSTION`` covers two raise sites in
``generate/stream.py`` — both express "destination-side admissibility
produced no admissible candidate":
1. Pre-walk: the region's allowed-index intersection with the
candidate set is empty before any step ran. ``step_index = -1``
and ``rejected_attempts = ()`` distinguish this site — no
inner-loop rejections were issued; the region was already empty.
2. In-walk: at some step, every candidate in the admissible set was
rejected by ``check_transition`` (threshold mode) or
``check_margin`` (Phase 3 / ADR-0026, margin mode). ``step_index
>= 0`` and ``rejected_attempts`` records the tried (index, word,
score) triples in attempt order (threshold) or the full ranking
(margin).
``ROTOR_REJECTION`` (Phase 4 / ADR-0025) is the rotor-side analogue.
Raised when no admissible destination produces a rotor that lands
the field within the region's frame versor cone. Carries the same
trace shape: ``rejected_attempts`` records the rotor-side scores
(``cga_inner(F_after, frame_versor)``) for each rejected candidate.
Splitting this from ``INNER_LOOP_EXHAUSTION`` keeps the trace
self-explanatory — a refused turn says *which* admissibility axis
ran out, destination-blade or rotor-frame.
"""
from __future__ import annotations
from enum import Enum, unique
@unique
class RefusalReason(Enum):
"""Machine-readable refusal taxonomy.
The string value is what flows into ``trace_hash`` payloads when
refusal materialisation is wired through a future ADR. Stable
string values are part of the replay contract.
"""
INNER_LOOP_EXHAUSTION = "inner_loop_exhaustion"
# Phase 4 / ADR-0025 — rotor side: every candidate's rotor would
# push the field outside the region's frame-versor admissible
# cone (``cga_inner(F_after, frame_versor) <= 0`` after
# sandwiching). Distinct from INNER_LOOP_EXHAUSTION so the
# trace can tell destination-blade refusal from rotor-frame
# refusal without re-parsing the message.
ROTOR_REJECTION = "rotor_rejection"
# W-011 — recognition-side refusals. The DerivedRecognizer
# refused the input (shape mismatch, missing feature evidence,
# or feature contradiction). Folded into CognitiveTurnResult
# .refusal_reason so the pipeline boundary does not discard the
# typed recognition refusal.
RECOGNITION_REFUSED = "recognition_refused"
class InnerLoopExhaustion(ValueError):
"""Honest refusal raised by the generation walk.
Subclasses ``ValueError`` so pre-Phase-2 ``except ValueError``
handlers in ``chat/runtime.py``, ``evals/...``, and tests still
catch it without modification.
Attributes
----------
reason : RefusalReason
Machine-readable taxonomy code. Phase 2 uses a single value;
Phase 4 (rotor frame admissibility) is expected to add more.
region_label : str
The ``AdmissibilityRegion.label`` that produced the refusal —
the operator-visible identifier of which constraint blocked
propagation.
step_index : int
Index of the step at which the inner loop exhausted. ``-1``
marks the pre-walk site (region intersection empty before any
step ran); non-negative values identify the in-walk site.
rejected_attempts : tuple[tuple[int, str, float], ...]
Ordered record of ``(candidate_index, word, score)`` triples
that the inner-loop check rejected at the step that failed.
Empty tuple for the pre-walk site.
``str(exc)`` returns the human-readable message — preserving the
Phase 1 exception message contract for callers that only inspected
``str(e)``. Instances are logically immutable: attributes are set
once in ``__init__`` and should not be reassigned.
"""
__slots__ = ("reason", "region_label", "step_index", "rejected_attempts")
def __init__(
self,
*,
reason: RefusalReason,
region_label: str,
step_index: int,
rejected_attempts: tuple[tuple[int, str, float], ...] = (),
message: str | None = None,
) -> None:
self.reason = reason
self.region_label = region_label
self.step_index = step_index
self.rejected_attempts = tuple(rejected_attempts)
super().__init__(message if message is not None else self._default_message())
def _default_message(self) -> str:
if self.step_index < 0:
return (
f"AdmissibilityRegion[{self.region_label}] left no walk candidates."
)
return (
f"AdmissibilityRegion[{self.region_label}] inner-loop "
f"rejected all candidates at step {self.step_index}."
)