feat(hslm): [I] Ternary 3-Phase LR Schedule — φ-decaying cycles

[gHashTag]

Parent: #329 — Fully Ternary Transformer

Level 0 (no dependencies) | ~60 LOC | New: src/hslm/ternary_schedule.zig

Description

3-phase learning rate schedule (warmup/cruise/cooldown) repeated 3 times with φ-decaying max LR: cycle 1 = max_lr, cycle 2 = max_lr × 0.618, cycle 3 = max_lr × 0.382.

Implementation

• TernarySchedule.getLR(step: u64, total: u64) → f32 — 3 phases × 3 cycles = 9 transitions
• getLRDecaying(step: u64, total: u64, max_lr: f32) → f32 — with φ^(-cycle) decay
• Phase boundaries: warmup 10%, cruise 60%, cooldown 30% per cycle

tri CLI

tri hslm lr-schedule

Tests

• 9 transitions: correct phase at each boundary
• LR bounds: never > max_lr, never < 0
• Decaying sum: cycle2_total < cycle1_total (by φ ratio)

Blocks

• Issue L (Integration — trainer.zig uses ternary schedule)

[gHashTag]

Consolidated into feat(hslm) epic #329 — Fully Ternary Transformer