🔍 debug(l6): expose claude -p output to CI logs (#116)

[ZaxShen]

Diagnostic PR for #116. Doesn't fix the L6 failure — gathers info to identify which layer breaks.

Problem (#116)

First L6 run on dev produced 0 trajectory rows + 0 tokens for all 4 wired flows. Runner suppressed claude's output via `>/dev/null` — we have no idea what claude was doing.

What this PR adds

1. Strip output suppression from 16 flow scripts (`l6_run_claude ... >/dev/null` → `l6_run_claude ...`)
2. Wrap claude output with `[claude]` prefix on CI's stderr
3. Pre-flight diagnostic steps in `run-l6.sh`:
   • `claude --version` (binary present?)
   • `claude -p "say hello in one word"` (auth works in -p mode?)
   • `claude --plugin-dir -p "say hi in one word"` (plugin loads in -p?)

What we expect to learn

After CI runs this on dev, the logs will show:

• If pre-flight ✨ v0.2.0 Phase 2 — SQLite MCP trajectory server #1 fails → `claude` not installed in CI
• If pre-flight 🐛 Phase 2.5 — fix 3 MCP correctness bugs from PR #1 review #2 fails → auth broken (token format / not recognized)
• If pre-flight 🐛 Phase 2.5 — fix 3 MCP correctness bugs from PR #1 review #2 succeeds + ✨ v0.2.0 Phase 3 — two-tier agent roster + skills ported to native layout #3 fails → `--plugin-dir` doesn't work in -p mode
• If both succeed but flows still fail → bro persona trigger not engaging in -p mode

Each finding has a different fix.

Cost

~$0 if any pre-flight fails (no completion). ~$0.30 if all 3 pre-flights succeed + 4 wired flows run.

After this

Based on the diagnostic output, follow-up PR (or comment on #116) to fix the broken layer. May lead to:

• L6 v3 with a different invocation pattern
• Falling back to L5+L6 combined Docker (which uses marketplace install path, not `--plugin-dir`)
• Documenting that L6 requires interactive mode

🤖 Generated with Claude Code