refactor: replace EMA-driven context cap with tier-based cost-aware decisions

[BYK]

Summary

• Remove the adaptive context cap system (EMA bust rate tracking, dynamic cap tightening/relaxation, MIN_CONTEXT_FLOOR, maxContextTokensCeiling) that caused death spirals by ratcheting the cap below the session's actual token count
• Replace with tier-based quality watermarks (200K/500K/model limit) and per-turn economic bust-vs-continue decisions using actual cache pricing data
• Add rolling bust detection: after 5 consecutive cache busts, stop compressing and inject a warning message advising the user to compact or start a new conversation

Key changes

New API

• setCachePricing(write, read) — sets per-token cache costs from models.dev
• shouldCompress(currentTokens, compressedTokens, consecutiveBusts) — bust-vs-continue economic decision with 0.85 threshold. Heavily favors NOT compressing since cache writes are 12.5x more expensive than reads on Opus. Returns false (don't compress) when no pricing data is available.
• getTier(tokens) — maps token count to quality tier (0: ≤200K best, 1: ≤500K acceptable, 2: >500K degraded)
• recordCacheUsage(write, read, inputTokens, sessionID) — tracks consecutive busts using total input tokens as denominator (not just cache tokens)

Integration points

• Tier gate in transformInner: Between the layer-0 passthrough and compression stages, shouldCompress() is called. When bust cost isn't justified, the session stays at layer 0 with full context up to the model limit.
• Unsustainable warning in pipeline.ts: When TransformResult.unsustainable is set (5+ consecutive busts), a <system-reminder> warning is injected into the last user message advising to compact or start fresh.
• DB persistence: consecutiveBusts is persisted via the repurposed dynamicContextCap column (no migration needed) and restored on session load.

Removed

• adaptContextCap(), computeContextCap(), setMaxContextTokens(), getMaxContextTokens(), updateBustRate()
• bustRateEMA, interBustIntervalEMA, lastBustAt, dynamicContextCap session state fields (DB columns retained, repurposed)
• effectiveCap constraint in transformInner — context now uses full model window
• targetBustCost, maxContextTokens config fields (marked deprecated, still parsed to avoid breaking existing .lore.json)

Why

The old EMA system optimized for minimizing input_tokens per turn via a static dollar-based cap. This created a death spiral: high bust rate → tighten cap → session can't fit in lower layers → forced to Layer 4 → bust rate stays high → cap ratchets down further. The tiers are quality-based (empirical model effectiveness), not pricing-based — token costs scale linearly regardless of tier.

Math verification (Opus 4.6)

• Cache write: $6.25/MTok, Cache read: $0.50/MTok (write is 12.5x more expensive)
• At 250K → 150K compressed: bustCost=$0.94, continueCost=$0.125 → don't compress ✓
• At 500K → 100K compressed: bustCost=$0.625, continueCost=$0.25 → don't compress ✓
• At 2M → 100K compressed: bustCost=$0.625, continueCost=$1.00 → compress ✓