simple-microchunk-rejection

name	simple-microchunk-rejection
description	Use when working on SIMPLE's streaming microchunk class-average rejection, simple_microchunked2D lifecycle sentinels, pass-1/pass-2/refchunk/match generation, skip_pass_2 behavior, simple_cluster2D_rejector criteria, or boundaries with model_cavgs_rejection.

SIMPLE Microchunk Rejection

Use this skill for src/main/stream/simple_microchunked2D.f90, src/main/stream/simple_cluster2D_rejector.f90, and stream integration questions involving model_cavgs_rejection.

Read First

• doc/microchunk_and_rejection/microchunk_rejection_policy.md
• doc/microchunk_and_rejection/microchunk_rejection_model_integration.md
• doc/microchunk_and_rejection/model_cavgs_rejection.md when comparing the rule engine with the learned model backend.

Live Ownership

simple_microchunked2D owns stream lifecycle, sentinel reconstruction, tier generation, match finalization, and combined outputs. simple_cluster2D_rejector owns the live scalar rejection rules. src/main/cavg_quality and model_cavgs_rejection are separate unless a task explicitly integrates them.

Sentinel Contract

Restart state is reconstructed from sentinel files:

• ABINITIO2D_FINISHED: chunk job finished.
• REJECTION_FINISHED: rejection and particle cleanup finished.
• REJECTION_FAILED: rejection could not be performed safely.
• COMPLETE: chunk was consumed or finalized.

On import, failed chunks count as rejection-complete and complete. Do not consume a source chunk into the next tier until its project file for the next tier has been written successfully.

Rejection Rules

The live rejector applies cumulative class-average rejection in this order:

1. population, with tier-specific fractions;
2. resolution, rejecting res > 40.0;
3. foreground mask geometry after low-pass/Otsu/connected-component cleanup;
4. local variance, including unconditional rejection of both-zero scores.

After rejection, rejected classes map to particle state=0, and rejected particles must have class=0 and class_match=0.

Tier Flow

The normal ladder is pass-1 -> pass-2 -> refchunk -> match. Pass-1 chunks become complete only after pass-2 projects are written. Pass-2 chunks feed the refchunk but are marked complete only after match chunks are generated. Match finalization updates accepted/rejected counters, writes COMPLETE, and copies project files to the completion directory.

When skip_pass_2 is active, pass-2 import and generation are skipped. Rejection-complete pass-1 chunks feed the refchunk directly, and remaining pass-1 chunks become match chunks directly after the reference stack and box are available.

Model Backend Boundary

model_cavgs_rejection is a learned feature-vector backend with hard validity gates, normalized features, model weights, clustering, and threshold controls. It currently exposes chunk_default_v2, chunk_lp4, pool_default_v2, microchunk_p1, and microchunk_p2 presets. Do not replace stream rejection decisions with model output unless the task also preserves stream sentinels, selected/rejected stacks, project state propagation, and match/finalization semantics.