Add governed QCoder implementation

[echoomegaprime]

Summary

Built the QCoder governed plugin and added a pinned Crucible security-tool catalog covering 20 additional open-source repositories across low, medium, high, and critical tiers. Added strict catalog validation, tests, release staging, and documentation. High-risk tools require an authorized Crucible scope; critical Caldera is Crucible-only.

Why

QCoder needs stronger coding, security validation, and autonomous-maintenance capabilities without turning a local terminal into unrestricted target or credential authority.

Validation

• pwsh -File .\scripts\verify-plugin.ps1 -SkipMcpSmoke — PASS
• npm run typecheck — PASS
• npm run lint — PASS
• npm test — PASS (server 34, web 7, script 16, golden 60/60)
• node .\scripts\validate-crucible-catalog.mjs — PASS (20; low 9, medium 5, high 5, critical 1)
• npm audit --audit-level=high — PASS (0 vulnerabilities)
• python .\launcher\tests\test_qcoder_adapter.py — PASS (17)

Security

Catalog entries are pinned to exact commits and license SHA-256 digests. Membership grants no shell, network, credential, or target authority. Low-risk scanners are workspace-safe; active discovery/fuzzing/proxy tools require Crucible scope; Caldera is critical and Crucible-only. No secrets or restricted data were added.

Evidence

Commit a8e4064 on agent/qcoder-canonical. Hosted checks must pass on this exact commit before merge.

[echoomegaprime]

Added batch 2 of the Crucible tool catalog (commit f00303e), requested as a follow-up after batch 1 above.

• Restructured config/crucible-tool-catalog.json to a multi-batch schema (schemaVersion: 2) — batch 1's 20 entries are preserved byte-for-byte.
• Batch 2 re-verified all 20 originally proposed candidates (Nmap, Masscan, Gobuster, Nikto, SQLMap, Amass, theHarvester, Recon-ng, Sherlock, Zeek, Suricata, Wazuh, Velociraptor, Falco, Bandit, Semgrep, CodeQL, TruffleHog, BrowserGym, Open Interpreter) live against GitHub. 12 failed the MIT/Apache-2.0 policy gate (custom/AGPL/GPL/unverifiable licenses) and were replaced with equivalent permissively-licensed tooling: Grant, OSV-SCALIBR, gosec, detect-secrets, Cosign, in-toto, OPA, CycloneDX CLI, honggfuzz, Garak, and PyRIT.
• Garak/PyRIT are the catalog's first generative-AI red-team entries; the validator's route check now also accepts echo.ai_redteam.* alongside echo.crucible.*.
• Every commit SHA and license SHA-256 was fetched live from the GitHub API; 4 repos whose GitHub SPDX classifier reported NOASSERTION (amass, grant, in-toto, BrowserGym) were confirmed genuinely Apache-2.0 by reading the actual LICENSE text before inclusion.

Validation: node scripts/validate-crucible-catalog.mjs — PASS (2 batches, 40 entries; batch 2 low=11 medium=6 high=2 critical=1). npm run verify (typecheck, lint, format, full test suite incl. server 34 / web 7 / script 16 / golden 60, build) and npm audit are green.

[echoomegaprime]

Added batch 3 of the Crucible tool catalog (commit 77c9024) — reverse-engineering and binary-analysis tooling, per an explicit follow-up request to cover the tool families named in FLEET_ROLES/reverse-engineer.md.

20 new entries, all live-verified against GitHub (commit SHA + license):

• Low (16): Ghidra, capa, FLOSS, YARA, jadx, Androguard, RetDec, LIEF, Capstone, Refinery, Detect It Easy, BinExport, plyara, CyberChef, PortEx, StringSifter.
• Medium (2): angr, pe-sieve.
• High (1): syzkaller.
• Critical (1): Frida — crucible-only, cataloged for capability discovery only (live cross-platform code injection into any process is the same open-ended blast radius as Caldera/Open Interpreter).

8 candidates were excluded after live verification confirmed copyleft licensing (MobSF GPL-3.0, Keystone GPL-2.0, Unicorn GPL-2.0, radare2 LGPLv3, rizin LGPL-3.0, Cutter GPL-3.0, Qiling GPL-2.0, objection GPL-3.0); x64dbg and pyelftools were left out rather than included unverified/out-of-policy.

Required two policy extensions (both documented in policy.note):

• allowedLicenses += BSD-3-Clause, BSD-2-Clause (Capstone/YARA/angr/Refinery/pe-sieve — GitHub's classifier reported NOASSERTION/None for several despite genuinely permissive BSD text, confirmed by reading the actual LICENSE files) and wxWindows-3.1 (Frida only).
• routePrefixes += echo.re., echo.mobile_pentest., echo.prometheus_strike. — real ECHO capability namespaces, live-confirmed via arcanum_sdk.sdk_capabilities rather than invented. Several batch-3 entries (ghidra→echo.re.binary.decompile_fn, jadx→echo.re.apk.decompile, floss→echo.re.binary.strings, capa→echo.prometheus_strike.capa_analyze, yara→echo.crucible.fw_yara, cyberchef→echo.crucible.cyberchef_serve, frida→echo.mobile_pentest.frida_attach) route to real registered capabilities rather than the echo.crucible.tool.* placeholder pattern.

The catalog test now derives its accepted route prefixes from policy.routePrefixes instead of a hardcoded regex, so this won't need a matching test edit on every future namespace addition.

Validation: node scripts/validate-crucible-catalog.mjs — PASS (3 batches, 60 entries; batch 3 low=16 medium=2 high=1 critical=1). npm run verify and npm audit are green.