Iter 7: Eval framework, architecture extraction, and tool behavior refactor

[echo-xiao]

Summary

• Architecture extraction: Added architecture.json and refactored registry.ts/retriever.ts for graph-native code analysis
• Eval framework: Added layered evaluation (layer0-baseline, layer1-tool, layer2-agent) with compare utility and Claude judge scoring
• Gemini answers: Updated all 27+ answer logs with improved formatting and refined content
• Cleanup: Renamed eval files to layer-based naming, added executive report, removed stale agent-eval.md

---

Executive Report

$(cat logs/executive-report-flash.md)

---

Layer 0 — Baseline Eval (No Tools)

$(cat logs/layer0-baseline-eval.md)

---

Layer 1 — Tool Eval

$(cat logs/layer1-tool-eval.md)

---

Layer 2 — Agent Eval (Summary)

Full report is too large for PR description. See logs/layer2-agent-eval.md for complete details.

Metric	Value
Good answers (3+ file paths)	22/34 (64.7%)
Weak answers (has content, <3 paths)	4/34
Empty answers	8/34
File hit rate (avg, string match)	33.8%
Symbol coverage (avg, string match)	63.5%
Avg tool calls / question	7.2
Avg tokens / question	31,295
Total tokens (all 34)	1,064,042

By Question Type

Type	Count	Passed	Rate
architecture	9	0	0.0%
call-chain	4	0	0.0%
pattern	6	0	0.0%
locate	8	1	12.5%
routing	4	1	25.0%
impact	3	0	0.0%

By Difficulty

Difficulty	Count	Passed	Rate
medium	17	1	5.9%
hard	17	1	5.9%

---

🤖 Generated with Claude Code

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.

---

Echo Xiao seems not to be a GitHub user. You need a GitHub account to be able to sign the CLA. If you have already a GitHub account, please add the email address used for this commit to your account.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}