Release clean DREL production candidate

[drxadqz]

Outcome

Publishes the clean, checkpoint-compatible DREL-E B component_full production candidate without copying the exploratory ablation framework or local experiment artifacts into the public repository.

What changed

• adds a self-contained DREL production module with:
  • one-way directional/radial evidence ledger
  • constrained matched-response filters
  • moment-preserving transitions
  • retained regional mean/deviation blocks
  • bounded 0.25 * tanh evidence writes
  • 320-D embedding and 27-way linear classifier
• preserves the research component_full state-dict schema
• adds a path-free frozen D350 specification and runnable forward/backward example
• adds five focused unit tests and extends the repository contract
• adds detailed English and Chinese beginner-facing algorithm guides
• adds machine-readable three-seed Gate8/Gate30 validation evidence and claim boundaries
• keeps the formally tested S7 result as the only headline test result

Why this version

The early Gate8 screen suggested deleting the regional decomposition, but the preregistered three-seed Gate30 deletion confirmation failed: seed297 reversed by -0.007 score and the mean Macro-F1 hurdle failed. RPCC/TERM, amplitude, late-amplitude, BLADE, and CORAL roughness routes also failed their frozen gates. The retained production architecture is therefore component_full, not the early-screen winner or an unfinished roughness experiment.

Evidence boundary

At equal D350 Gate8 validation budget, the DREL three-seed mean is above the frozen single-seed RSPNet-M/L envelope on 8/9 tracked metrics. Roughness is the exception: 0.586420 versus 0.608889 (-2.247pp). Gate30 roughness reaches 0.622963, but there is no equal-budget RSPNet Gate30 reference. No DREL development proxy-test or historical 49,500-image formal test was accessed or used for model selection.

Verification

• research-vs-release backbone: 2,751,694 parameters, identical ordered state-dict keys, identical tensors, bitwise-identical forward output
• strict load of the real seed297 Gate30 component_full checkpoint: all 330 state tensors matched
• complete classifier parameter count: 2,760,361
• python scripts/check_repository_contract.py
• python -m compileall -q src examples tests scripts/check_repository_contract.py
• python -m pytest -q tests/test_drel.py -> 5 passed
• git diff --check

Review notes

This is intentionally a Draft PR. The active RFCE roughness screen remains outside this release until it passes its preregistered Gate3/Gate8 decision rule.