agent_hint: per-agent split + codex/gemini opener tune

[Prasanna721]

Summary

• Split render_block into per-agent dispatcher with render_claude_block (Claude) and render_default_block (Codex + Gemini). Bodies are byte-identical except the opener.
• Tweak the codex/gemini opener to is searchable by meaning, not just filename. Codex was spiraling on filename-based filesystem exploration without an explicit anti-pattern phrase; this nudges it toward smfs grep as the primary search.
• Collapse the verbose 3-line smfs grep banner into a single line. Same info, less noise per query.

Measured impact

On a 10-question, 20-record codex/smfs subset vs codex/fs baseline:

state	codex pass	codex tok/correct	vs codex/fs (896,630)
pre-split (single body, claude-opener)	18/20	488,263	-45.5%
post-tune (codex opener)	18/20	420,273	-53.1%

The opener tune alone is -13.9% codex tok/correct at constant pass rate (26 fewer tool calls, 130s faster wall time). Claude's number was already optimal at 326K with the older opener, so the claude block was left alone.

Test plan

• cargo build --release -p smfs-core -p smfs clean
• cargo test -p smfs-core (note: there is a pre-existing stale test on a sibling branch asserting the older hint wording — not included here, so this branch's tests should pass)
• Manual: smfs mount <tag>, verify ~/.claude/CLAUDE.md and ~/.codex/AGENTS.md get different openers; verify both unmount cleanly
• Sanity sweep on a small dataset to confirm codex/smfs token usage stays in the same ballpark