fix(model-router): add claude-sonnet-5 catalog entry; classify recon subagents as light

[pimmink]

Problem

Two catalog gaps discovered in v1.12.0 during production configuration work:

1. claude-sonnet-5 absent from MODEL_CAPABILITY_TIER

GitHub Copilot, Anthropic direct, Vertex, and Bedrock all serve claude-sonnet-5 as GA.
isKnownModel("claude-sonnet-5") returns false because the model has no entry in
MODEL_CAPABILITY_TIER. Since #2192, an unknown configured model causes
resolveModelForComplexity to short-circuit and return that model unconditionally —
routing is silently bypassed for every task in that phase.

Users who set models.<phase>: claude-sonnet-5 in PREFERENCES.md lose dynamic
downgrade to haiku for light tasks and lose the escalation chain entirely.
tier_models.standard: claude-sonnet-5 works correctly (used as dispatch string),
but the models.* ceiling path does not.

The model is already registered in claude-code-cli/models.js — this is a catalog-only
omission in model-router.ts.

2. Recon subagents (codebase-mapper, scout) dispatched at standard tier

All subagent unit types default to "standard" via the UNIT_TYPE_TIERS[unitType] ?? "standard" fallback in complexity-classifier.ts. Read-only filesystem scanning agents have no reasoning requirement and should dispatch to haiku.

Changes

• model-router.ts: adds claude-sonnet-5 to MODEL_CAPABILITY_TIER (standard), MODEL_COST_PER_1K_INPUT, and MODEL_CAPABILITY_PROFILES
• complexity-classifier.ts: adds subagent/codebase-mapper and subagent/scout to UNIT_TYPE_TIERS as light
• tests/model-router.test.ts: regression tests for sonnet-5 catalog presence and routing-not-bypassed behaviour; adds MODEL_CAPABILITY_TIER to imports
• tests/complexity-classifier.test.ts: regression tests for subagent light tier

Non-breaking

Zero deletions. No changes to the #2192 bail-out logic. Existing tests unchanged.
subagent/security-auditor, subagent/code-reviewer, and all other analysis agents remain at the "standard" default.

Capability profile rationale for claude-sonnet-5

Scores are conservative, derived from published Anthropic benchmarks:

Dimension	Sonnet 4.6	Sonnet 5	Delta	Rationale
coding	85	90	+5	SWE-bench improvement
debugging	80	85	+5	Correlated with coding
research	75	80	+5	Improved tool use
reasoning	80	87	+7	Most notable improvement
speed	60	55	-5	Slightly larger model
longContext	75	80	+5	1M context window
instruction	85	88	+3	More instruction-faithful

Closes #1612

[github-actions]

🟡 PR Risk Report — MEDIUM

Files changed	4
Systems affected	1
Overall risk	🟡 MEDIUM

Affected Systems

Risk	System
🟡 medium	Model System

File Breakdown

Risk	File	Systems
🟡	src/resources/extensions/gsd/model-router.ts	Model System
⚪	src/resources/extensions/gsd/complexity-classifier.ts	(unclassified)
⚪	src/resources/extensions/gsd/tests/complexity-classifier.test.ts	(unclassified)
⚪	src/resources/extensions/gsd/tests/model-router.test.ts	(unclassified)