LEDGER VERDICT: GO (docs/language-dichotomy-ledger.md) — the journal
write is I/O-bound file IO (decision tree: I/O-bound → Go) and the
census is simple counting over a static table (Laws 3/7: simple
operations stay in the caller's language; FFI overhead would exceed
the compute benefit — the census_report fold is ~16 increments).
parity: every public function of inference_paths ported — the CENSUS
table (all 16 call sites, insertion order preserved, notes verbatim),
trace_side_path's record shape (schema_version 1, kind side_path,
module/function/model/task_id/duration_ms, at) with the no-clock law
(`at` is a kernel argument — the oracle renders it via
time.strftime('%Y-%m-%dT%H:%M:%SZ', time.gmtime()); the freeze pins
time.gmtime), and census_report (total / by_kind fold / the EXCLUDED
list with documented reasons). Boundary law: the JOURNAL attempt
(rune.inference_trace import + record_trace + the journaled flag)
stays in the Python facade — the Go kernel owns the record shape and
the report. ORACLE DIVERGENCE NOTE (observable behavior pinned, not
fixed): the oracle's journal path is currently dead — InferenceTrace's
signature drifted (requires resolved_context_window/input_budget/
model_calls; the oracle passes prompt_tokens/output_tokens/cost), so
trace_side_path ALWAYS returns journaled=False via the except branch.
The frozen truth records that faithfully; the facade keeps its own
float duration_ms (Python renders 0.0 where Go marshals 0 — the
battery compares semantically). Journal attempts in the battery write
to throwaway temp dirs, never the real journal. gen_inference_paths_
truth.py freezes 7 cases (record defaults/full/zero/float/empty +
journal-failure branch + full census report); replay mode re-verifies
the live oracle 7/7. Go battery: go test replays truth.json VERBATIM +
the `inference-paths replay` CLI — 7/7; gofmt clean, go vet clean;
test_inference_paths.py 6/6 (no Python-side delegation pin — Go ports
are standalone per recall-uplift precedent).
