feat(mcp): --with-output-style + stop-condition CLAUDE.md additions (banger phase 16)

[aomerk]

Summary

• Ships keeba mcp install --with-output-style — drops ~/.claude/output-styles/keeba.md, a terse engineering style that suppresses preamble + tool-result restatement + closing summaries (the three biggest output-token sinks in a typical Claude Code investigation). Activate per-session with /output-style keeba.
• Strengthens the --with-claude-md template with an Answer discipline (output-token savings) subsection: one-find_def-hit-stops, quote-don't-restate, no-preamble, no-closing-summary, conclusion-first.
• README install snippet now leads with all four UX patches (--patch-agents --with-claude-md --with-hook --with-output-style).

Why

Real-world A/B (Slack-thread investigation prompt) showed our existing codec layers ceiling at ~30% session-cost reduction:

arm	cost	cache_read	output
no-keeba	$3.73	2.6M	26.2k
keeba (lean codec L2)	$2.73	1.3M	19.2k

L1 (hook codec) + L2 (lean MCP responses) both shrink cache_read (priced ~$0.30/M). But session cost is dominated by output tokens (priced ~$15/M — 50× more) and output didn't shrink. To break the ceiling we have to attack the output side.

Output tokens come from five places. We control three:

source	controllable?	lever
Preamble ("I'll investigate by...")	yes	output-style: No preamble directive
Tool-use args	no (small)	—
Restatement of tool results	yes	output-style + CLAUDE.md: Quote, don't restate
Wandering / "let me also check" reasoning	yes	CLAUDE.md Answer discipline stop conditions
Internal thinking blocks	no (Opus extended thinking, user-toggled)	—

This PR ships the levers for the three controllable sources. They are opt-in (existing installs unaffected) and idempotent (re-running keeba mcp install prints "no change").

Why opt-in, not default

Output styles in Claude Code replace the default engineering system prompt. Mutating the user's defaultOutputStyle from an install is too invasive — they may have their own style or rely on the default. Drop the file, tell the user to /output-style keeba. They opt in.

What's NOT in this PR

• Lever 4 (vertical agent — keeba investigate runs the loop in our control plane and returns one finished answer for Claude to synthesize). That's a much bigger build; defer until we measure whether output-style + Answer discipline already break the ceiling.
• Mutation of ~/.claude/settings.json defaultOutputStyle. Opt-in only.

Test plan

• go test ./... — all green
• gofumpt -l clean on changed files
• go vet ./... clean
• golangci-lint run ./internal/cli/... — 0 issues
• End-to-end install dry run against a temp HOME:
  • --with-output-style lands ~/.claude/output-styles/keeba.md with valid frontmatter (name: keeba)
  • --with-claude-md injects ## Code investigation in keeba-indexed repos with the new Answer discipline subsection containing NO preamble, NO closing summary, Quote tool-result rows verbatim, 50× cache_read
  • re-run prints "no change" for both
• Verification A/B (post-merge, run by user): install with --with-output-style, activate /output-style keeba, re-run the same Slack-thread prompt, compare cost vs the $2.73 lean-codec baseline. Expected: <$2.00 if the lever is real; ~$2.73 if output bloat is structural and we need lever 4.

🤖 Generated with Claude Code