SD-228

[layer-model-v03] Layer model bumped to v0.3 — 18 evidence annotations from empirical data. Model bump workflow per SD-220: archived v0.2 to docs/internal/archive/layer-model-v0.2.md (444), applied 18 annotations/additions to docs/lexical-harness-not-prompt-harness.md, bumped version. Constraint: annotations and additions ONLY, no structural mutations — Captain’s ruling (SD-220). Sources: (1) layer-model-annotations-sd195-sd203.md — 4 SUPPORTS, 4 CHALLENGES, 5 ENRICHES. (2) layer-model-v03-suggestions-post-compaction.md — 5 post-compaction observations. (3) prospective-metacognitive-regulation-analysis.md — L12 trained capacity recommendation. Key additions by layer: L3 — recovery asymmetry primitive, phase transition concept, file consolidation evidence. L5 — compaction calibration enrichment. L6 — L6d BYPASS mode (file-mediated state outside harness, .keel-state), “mediates ALL” amended to “mediates all harness-mediated” with bypass noted. Captain ruled L6d contained within L6 structure (“not especially granular about L6, most is black box”). L7 — git as audit channel. L8 — stale entry evidence, named convention evidence, saturation threshold (arXiv:2602.11988). L9 — “monotonically increasing” caveat, partial recovery anchoring, git position trail. L10 — degenerate N=1 ensemble warning. L12 — context pollution evidence, terminal HUD instrument, backup storage (SD-205), trained capacity (Fonagy & Luyten, 2009). Also: agent-boot-audit added to holding deck — audit all agent files for boot context completeness and reference map gaps. Gate green (1,115 tests).