- Replace archived iter16 parquets as noise-floor source with a
  dedicated iter16b re-run (prior framing measured prompt drift from
  c1ec175, not same-prompt variance).
- Rewrite FAIL path to reflect that annotator emits all det_* columns
  in a single call, so any prompt iteration requires a full iter16-side
  rerun (~$15); only the report filters to regressed labels.
- Specify fingerprint-based row alignment against input parquet
  (non-det_* columns), since 5065 rows contain only 5059 unique text
  values.
- Drop "noise floor informational only" framing; noise floor now
  formally bounds override eligibility via |Δagr| ≤ 2 × noise_floor.
- Cap annotation concurrency at 2 jobs in flight (40 LLM requests)
  instead of running all sides in parallel.
- Type hints use pl.DataFrame, not pd.DataFrame, to match the
  existing polars-only audit stack.
- Expand CLI globs internally via Python glob, not the shell; assert
  each glob resolves to exactly 3 parquets.
- Total annotation cost: ~$45 (iter15 ×1 + iter16 ×2).

- FAIL-path iteration reruns both iter16a AND iter16b (~$30), not
  just iter16a. Reusing cached iter16b against a new iter16a prompt
  would turn the "noise floor" back into prompt-drift measurement —
  the exact mistake the Codex review caught for archived parquets,
  repeated at a different layer.
- Stratified-row filter uses audit_source == "stratified"
  explicitly; base cohort gets that value and injected rows use
  inject_<label>, so the original audit_source=injected phrasing
  would have included injected rows in agreement/prevalence metrics.
- prev_ratio 0/0 is now explicitly defined as 1.0 (no WARN), with
  a table for the other zero combinations. A silent-on-both-sides
  label no longer trips WARN-prevalence.
- CLI glob validation now parses filenames and asserts the model
  slug set matches the canonical {gemini3flash, sonnet46, gpt54mini}
  set rather than just counting 3 parquets. Catches duplicates and
  misnamed files that would silently corrupt agreement estimates.
- after and noise_floor det_* column sets are hard-asserted equal,
  not "assumed equal by construction." Override eligibility depends
  on noise floor so a schema mismatch must fail loud.
- Added unit test cases for each new validation.

Comprehensive task-by-task TDD plan for the iter17 A/B regression
audit. 16 tasks across 6 phases (P0 ops fixes → P1 audit helpers →
P2 compare_prompt_versions.py module → P3 annotation runs → P4
report generation → P5 gate decision).

Also aligns spec model slug to existing 'sonnet' convention
(not 'sonnet46'), matching the filename pattern of the existing
iter16_5k_sonnet.parquet artifact.

Restores python -m trainr.core.annotate_detections as a working
invocation path. iter16 lost ~30 min diagnosing a silent exit 0
when running via python -m; the click CLI was always the canonical
path but the backup diagnostic route was broken.

iter16 found 9 of 9 stack_trace 'misses' were actually regex
overmatches on Rust test directives, prose with 'at line N', and
Java frames without Exception headers. Tighten:

- Python: require 'Traceback (...)' header + 'File "' adjacency
- Java:   require 'Exception in thread' header + java-frame adjacency
- Rust:   (?m)^ line anchor excludes '// error-pattern:' directives

Patterns use [\s\S]{0,N} instead of (?s:.{0,N}) and (?m)^ instead
of negative lookbehind, because polars uses Rust's regex crate
which rejects both of the latter constructs.

Adds regression tests in both directions — previously-matching
anti-cases no longer match, real traces still match. Does not
rebuild iter16_5k_input.parquet.

Per-label majority-of-N fire rate on stratified rows. Needed by
the iter17 A/B compare module's soft WARN-prevalence gate. Uses
the same dict[model_slug, DataFrame] convention as the existing
compute_agreement_across_models. Filters internally so callers
don't need to remember to pre-filter to stratified rows.

Adds a test for the "label in first frame but not others" case
that exercises the continue branch in compute_prevalence_per_label.
Was flagged as a minor coverage gap in code review of 8f1ef7d.

Parses model slug from filename and asserts the parsed set equals
the canonical {gemini3flash, sonnet, gpt54mini} set. Rejects wrong
count, unparseable names, duplicates, and unknown slugs. Shared by
both the existing audit_semantic_labels.main() and the upcoming
compare_prompt_versions module.

The EXPECTED_MODEL_SLUGS frozenset is the single source of truth
for the canonical slug set.

Adds a test where two paths in different directories both parse
to the same canonical slug (gemini3flash). Exercises the
"if slug in parsed" branch in load_annotator_parquets that was
flagged as a minor coverage gap in review of 99159b4.

Regenerates the iter16 audit report from the archived parquets
and diffs against the committed docs/accuracy_runs report. Fails
loud on any silent drift in agreement math, rounding, or output
format. Insurance for the Phase 1 refactor + any future edit.

Defines LabelCategory, LabelVerdict, LabelRow, and DeltaReport
dataclasses. The compare_prompt_versions() function handles the
all-shared happy-path case with placeholder PASS verdicts.
_compute_prev_ratio implements the spec's zero-handling rules
(0/0 → 1.0, 0/>0 → inf, >0/0 → 0.0).

Verdict logic, column categorization, fingerprint validation, and
report emission arrive in subsequent tasks.

Dynamic det_* introspection categorizes every label as
shared/iter15-only/iter16-only. Shared labels get full metrics;
the other two categories get partial metrics and NO_VERDICT.
Hard-asserts after and noise_floor have identical det_* column
sets — the override criterion depends on noise floor, so a
schema mismatch must fail loud rather than silently skip.

Implements _compute_verdict() combining the hard agreement gate
(|Δagr| > 0.005) with the soft prevalence gate (prev_ratio outside
[0.5, 2.0]). Verdicts can co-occur (FAIL_AND_WARN). Includes
tests for each path including the zero-prevalence edge cases.

Override eligibility (|Δagr| ≤ 2 × noise_floor) is NOT applied
here. The verdict reflects the raw gate outcome; human review
applies override logic per the Gate Decision Protocol.

Validates every annotation frame against the input parquet by
comparing all non-det_* columns row-for-row. Raises ValueError
naming the first diverging row index, column, and source slug
on mismatch.

Needed because iter16_5k_input.parquet has 5065 rows but only
5059 unique 'text' values — text-only equality is insufficient
as an alignment check.

Fingerprint runs BEFORE the schema assertion so row drift is
caught with a precise row index rather than a vague column-set
diff.

format_delta_report renders the DeltaReport with all sections
from the spec: gate verdict summary, shared labels table,
iter15-only and iter16-only sections, FAIL-agreement detail
with override-eligibility hint (|Δagr|/noise_floor ratio),
WARN-prevalence detail, and a noise floor table covering all
categorized labels.

Wires compare_prompt_versions into the trainr data click group.
Module main() parses argv, resolves globs via Python glob module
(not the shell), loads parquets via load_annotator_parquets with
slug validation, runs the comparison, and writes the markdown
report. Exits 2 on any FAIL-agreement verdict so shell callers
can gate on it.

Code work (Phases 0-2, 11 tasks) is done and tested. Remaining
work (Phases 3-5, ~$45 LLM budget + human gate review) picked up
in a future session. Handoff doc captures:

- Completed task list with commit SHAs
- Test coverage summary (24 compare tests + 14 audit + 6 build_audit)
- Exact commands to resume at Task 3.1 (worktree setup)
- Gate decision protocol reference
- Pyright noise context + 10 unrelated pre-existing failures note
- FAIL-path cost note ($30 per iteration, iter15 cached)