# RECENT VIOLET 34E — trade/slot-granularity comparison harness

This pass implemented the next numbered comparison item from the main Violet spec:
the trade/slot-granularity comparison that was still deferred after the V3.4c
live-source work and shadow journal wiring.

## What existed before

Before this pass, Violet already had:

- `VioletDecisionEngine` producing shadow decisions
- `VioletDecisionJournal` persisting actuated decisions to `violet_decisions`
- live-source adapters for BLUE-published factors
- full sizing parity through `VioletSizer`

What was still missing was a dedicated comparison unit that could collapse those
journaled decisions into episode-sized runs and compare them against the terminal
trade surface at the same granularity.

## What was added

I added a new Violet-only comparator:

- [prod/clean_arch/violet/trade_slot_compare.py](/mnt/dolphinng5_predict/prod/clean_arch/violet/trade_slot_compare.py)
- [prod/clean_arch/violet/test_violet_trade_slot_compare.py](/mnt/dolphinng5_predict/prod/clean_arch/violet/test_violet_trade_slot_compare.py)

The new module provides:

- `DecisionEpisode`: a collapsed run of contiguous shadow decisions for one
  asset/side
- `TradeEpisode`: a collapsed terminal trade record grouped by `trade_id`
- `EpisodeMatch`: a paired decision/trade episode with timing and row-count gaps
- `TradeSlotComparison`: the full comparison result
- `compare_trade_slot_granularity(...)`: the top-level collapse-and-compare API

## How it works

The comparator is deliberately narrow:

- it groups decision rows by `asset` + `side`
- it splits them into episodes when the scan number gap exceeds the configured
  `max_scan_gap`
- it groups trade rows by `trade_id`
- it ignores malformed rows rather than failing on them
- it matches episodes by asset/side and preserves unmatched decision/trade rows

This is downstream of the existing shadow journal. It does not reimplement ACB,
EsoF, OB, or sizing math. It only compares the surfaces that already exist.

## Anomaly handling

The comparator rejects or skips bad inputs instead of letting them pollute the
episode view:

- missing asset or side
- missing or invalid scan number
- malformed or missing trade id
- non-finite numeric fields

That keeps the harness usable against noisy journal extracts and historical trade
rows that may contain replay artifacts.

## Tests

The new tests cover:

- episode collapse across contiguous decision rows
- terminal trade dedupe through `trade_id`
- scan-gap splitting
- malformed row handling
- empty-input behavior

Verification on this pass:

- `PYTHONPATH=/mnt/dolphinng5_predict rtk python -m pytest -q /mnt/dolphinng5_predict/prod/clean_arch/violet/test_violet_trade_slot_compare.py`
- result: `3 passed`

## Why this is the right next step

The main Violet plan explicitly defers trade/slot-granularity comparison until the
shadow side has a comparable execution surface. That condition is now met well
enough to build the comparison harness without touching BLUE or the live executor.