core/workbench/replay.py
Shay 53ff9359a8 feat(workbench): Wave M C3 field substrate — persist-first field evidence
Makes the CL(4,1) field geometry legible, honestly. The engine owns the field;
this surfaces ONLY the exact scalar invariants it computes — never the raw
multivector — so the workbench shows "this is the geometry, it's exact, it
can't fake coherence" without a decorative blob or any motion.

Persist-first (the C3 gating work was Python, not React): the honest scalars
were computed live per turn but discarded. Now captured per turn into the
journal so a read-only surface has real evidence.

Backend:
  - workbench/field_evidence.py: FieldEvidence computed from the engine result —
    exact versor_condition, field_valid (vs the 1e-6 ceiling), a content-
    addressed field_digest (sha256 of the engine-canonical array bytes), and
    cga_inner(before, after) as the exact transition value. Raw field bytes
    never cross the boundary: only floats + digests. validate() is fail-closed —
    field_valid can never disagree with versor_condition vs the ceiling (the
    wrong=0 analogue for the geometry). No engine math re-implemented
    (versor_condition / cga_inner imported from algebra; bytes via array_codec).
  - workbench/schemas.py: FieldEvidence dataclass; field_evidence on
    ChatTurnResult + TurnJournalEntrySchema. schema-snapshot.json regenerated.
  - workbench/journal.py + api.py: persisted at from_chat_turn; first-class read
    endpoint GET /trace/{turn_id}/field (trace facet, consistent with /pipeline).
  - workbench/replay.py: field_evidence classified CRITICAL — replay now also
    proves field determinism (digest + scalars must match on re-execution).

Frontend:
  - types/api.ts FieldEvidence + field_evidence passthrough; client/query hook;
    FieldInvariantCard (measured value vs ceiling, cga_inner transition, digests;
    honest missing_evidence; no blob, no motion); Trace route Field tab.

Honest-empty for pre-widening journal rows (missing_evidence). Deferred:
cross-turn field-coherence trends, session-level field persistence.

Validation: 138 workbench/practice Python tests (incl. non-vacuous field guards
+ replay field-determinism); 465/465 frontend incl. schemaDrift; pnpm build
clean; git diff --check clean. No generate.derivation / reliability_gate /
stream / field.propagate / vault.store imports.
2026-06-13 16:26:46 -07:00

111 lines
4 KiB
Python

"""Sealed single-turn replay over the turn journal (Wave R3).
``GET /replay/{turn_id}`` re-executes a journaled prompt in a sealed fresh
runtime and compares the resulting envelope leaf-by-leaf against the
recorded :class:`~workbench.journal.TurnJournalEntry`.
The claim demonstrated is exactly the architectural one: same prompt, same
genesis substrate -> bit-identical envelope. The original turn ran in its
own fresh ``ChatRuntime()`` that may have loaded an engine-state checkpoint
present at the time; the journal does not record whether one existed, so
``origin_state`` is reported as ``"unrecorded"`` and a divergence means
nondeterminism OR origin-state influence — never claimed to be one or the
other.
This module owns classification and comparison only. Execution is
injected by the caller (``workbench.api`` passes its own chat-turn
executor over a sealed runtime), which keeps the comparison pure and the
no-fabricated-equivalence obligation testable: if the executor raises,
no comparison object exists.
"""
from __future__ import annotations
import time
from dataclasses import fields, replace
from typing import Callable
from workbench.journal import TurnJournalEntry
from workbench.schemas import (
ChatTurnResult,
TurnReplayComparison,
TurnReplayDivergence,
)
# Every TurnJournalEntry field must appear in exactly one of these sets —
# enforced by tests so a future journal field forces an explicit
# classification decision instead of silently defaulting.
CRITICAL_FIELDS = frozenset(
{
"turn_id",
"prompt",
"surface",
"articulation_surface",
"walk_surface",
"trace_hash",
"grounding_source",
"epistemic_state",
"normative_clearance",
"verdicts",
"refusal_emitted",
"hedge_injected",
"proposal_candidates",
"leeway_evidence",
"pipeline_record",
"field_evidence",
"checkpoint_emitted",
"trace_integrity",
}
)
# Wall-clock by nature, or derived over wall-clock bytes (journal_digest
# hashes the timestamp): expected to differ on every replay and never
# evidence against equivalence.
INFORMATIONAL_FIELDS = frozenset({"timestamp", "turn_cost_ms", "journal_digest"})
def replay_turn(
entry: TurnJournalEntry,
*,
execute: Callable[[str], ChatTurnResult],
) -> TurnReplayComparison:
"""Re-execute ``entry.prompt`` via ``execute`` and compare envelopes.
``execute`` must run the prompt through the same envelope-assembly path
that produced the recorded entry, over a sealed runtime. This function
never fabricates a comparison: if ``execute`` raises, the exception
propagates and no ``TurnReplayComparison`` exists.
"""
started = time.perf_counter()
result = execute(entry.prompt)
elapsed_ms = max(0, int(round((time.perf_counter() - started) * 1000)))
result = replace(result, turn_cost_ms=elapsed_ms, turn_id=entry.turn_id)
replayed = TurnJournalEntry.from_chat_turn(result, turn_id=entry.turn_id)
divergences: list[TurnReplayDivergence] = []
for spec in fields(TurnJournalEntry):
original_value = getattr(entry, spec.name)
replay_value = getattr(replayed, spec.name)
if original_value == replay_value:
continue
divergences.append(
TurnReplayDivergence(
path=spec.name,
original=original_value,
replay=replay_value,
severity=(
"critical" if spec.name in CRITICAL_FIELDS else "informational"
),
)
)
return TurnReplayComparison(
turn_id=entry.turn_id,
comparison_basis="sealed_fresh_runtime_single_turn",
origin_state="unrecorded",
original_trace_hash=entry.trace_hash,
replay_trace_hash=replayed.trace_hash,
equivalent=not any(d.severity == "critical" for d in divergences),
replay_turn_cost_ms=elapsed_ms,
divergences=divergences,
leeway_evidence=entry.leeway_evidence,
)