Reduce subagent token usage across the forge skill chain

[anthonyespirat]

Summary

Implements the token-usage improvements drafted in webapp/improvement.md (scratch notes, not tracked). Net: -49 lines, 4 files — and a much tighter per-dispatch context profile. The changes compound: every dispatch reads a shorter prompt, returns a shorter report, and defaults to a cheaper model unless the step actively asks for opus.

What changed

• skills/subagent-execution/implementer-prompt.md (156 → 81 lines). Dropped the duplicated debugger / in-over-your-head / code-organization prose — the rules already live in skills/debugger/, so a one-line pointer is enough. Collapsed Forbidden to two bullets. Capped the report at 150 words and made the IF BLOCKED block conditional, so DONE reports stop dragging it back into the controller's context. Added an EXCERPTS-first instruction so the subagent skips redundant Reads when signatures are pasted inline. Removed the redundant Status meanings section — the controller's handling narrative in SKILL.md is the one that matters.
• skills/subagent-execution/SKILL.md. Inverted the model default: sonnet is now the default, opus requires an explicit upgrade signal (3+ files across modules, design judgment, BLOCKED reasoning re-dispatch). Added a Mechanical steps — skip dispatch section so the controller can execute fully-specified [mechanical] steps directly instead of paying for a round-trip, and tracks them as DIRECT EXECUTIONS in the final report.
• skills/writing-plans/SKILL.md. Added a [mechanical] step tag, a per-step EXCERPTS: block, and a ≤6-files-per-step budget. Extended the self-review with matching mechanical-audit and file-count-budget checks.
• agents/codebase-explorer.md. Flipped caveman compression from optional to default-on. The report is consumed by writing-plans, not a human, so the ~50% token cut compounds across every downstream dispatch.

Expected impact (per 10-step run, estimated)

Change	Savings
Trimmed implementer-prompt.md	~40–60k tokens
sonnet default + opus-on-signal	~30–40% on matching steps
Skip dispatch for [mechanical] steps	2–3 entire dispatches
150-word report cap, IF BLOCKED conditional	~10–15k tokens
EXCERPTS blocks cutting re-confirmation Reads	~5–10k tokens