chore: containerized dev env, Scorecard hash-pinning, agent-docs audit fixes

[dekobon]

Three related chores: a containerized dev environment, OpenSSF Scorecard remediation, and an audit-driven cleanup of the agent instructions that both efforts surfaced.

chore(docker): containerized dev environment for Claude Code + MCP servers (f7afe10)

Self-contained dev container that runs Claude Code plus the project's MCP servers (Context7, codegraph, Serena) and the full make pre-commit / make ci gate against a mounted checkout.

• Dockerfile (Ubuntu noble): Rust 1.94 stable + nightly, Node 24, Python 3.12 + uv/maturin, rust-analyzer/pyright/clangd/bash language servers, and the lint tools pinned to the CI versions (SHA256-verified).
• MCP servers are registered into Claude Code's user scope at build time from docker/mcp.json, so claude launches them as stdio subprocesses — no docker-compose, no separate server containers.
• Makefile targets dev-env-build/run/shell/rm (plain docker build/run; host UID/GID passed as build args so the bind mount stays writable).
• Optional docker/certs/ drop-in trusts corporate/proxy root CAs for building behind a TLS-intercepting proxy; ships empty, certs git-ignored.

Verified building and running on linux/arm64.

chore(security): hash-pin CI Python tooling, clear RUSTSEC advisories (154e332)

Addresses three OpenSSF Scorecard alerts:

• Pinned-Dependencies: CI workflows now install Python tooling from hash-pinned exports of uv.lock (big-code-analysis-py/requirements/{dev,examples}.txt, regenerated by make py-relock) with pip install --require-hashes, replacing floor-range installs and unpinned pip install --upgrade pip steps. Wheel smoke jobs install the just-built wheel by explicit dist/*.whl path with --no-deps. This also closes the "CI does not consume uv.lock" gap.
• Vulnerabilities: bump transitive deps anyhow 1.0.102 → 1.0.103 (RUSTSEC-2026-0190) and memmap2 0.9.10 → 0.9.11 (RUSTSEC-2026-0186); neither affected API is called directly.
• Binary-Artifacts: remove test_ext, a 4.3 MB compiled debug binary accidentally committed at the repo root. Nothing referenced it.

docs(agents): fix stale references across skills and instructions (9f7a396)

Audit of CLAUDE.md, AGENTS.md, and all 17 .claude/skills/ definitions against the repo surfaced ~35 verified defects, all fixed. Highlights:

• Default branch corrected from master to main in CLAUDE.md and seven skills whose git checkout -b / git diff commands were failing.
• Four skills invoked the retired CLI form big-code-analysis-cli -m -O json -p <path> (silently no-opping behind || true); now bca metrics -O json <path>.
• AGENTS.md still described a published 1.x library with breaks reserved for 2.0; reframed to the current major line / next major (3.0). Stale macro paths fixed.
• MCP dependencies degrade gracefully: skills no longer hard-require the unconfigured sequential-thinking server or a nonexistent serena:activate_project tool.
• issue-triage no longer creates GitHub labels despite declaring itself read-only.
• improve-crate, cleanup-crate, and batch-fix now run the canonical make pre-commit gate on integration branches; batch-fix agents gained the snapshot-anchoring and output-submodule discipline.
• scan-project partition fixes (c, mozcpp, objc, irules were in no partition), plus assorted smaller corrections (nexits vs exit, dead allowed-tools entries, missing workspace crates in listings).

Validation

• make pre-commit clean on the branch (cargo fmt/clippy/test/doc, udeps, lint families, man-page drift, self-scan gates, Python stages).
• Dev container image build + interactive run verified on linux/arm64.
• CI workflow edits validated with make actionlint.