feat: backtrader-agent P0 runtime + reproducibility hardening

[cloudQuant]

Summary

Introduces the backtrader-agent P0 runtime and a hardening pass that makes the
documented acceptance evidence reproducible in a fresh checkout and in CI. Before
this PR the repo held only .gitignore + LICENSE; the product had never entered
git, and the IMPLEMENTATION_REPORT's "59 passed / acceptance passed" claims did not
reproduce (15 tests failed out of the box).

What's included

Layered, offline-first, deterministic Backtrader strategy-authoring runtime:

• Python runtime (src/backtrader_agent): typed CLI, opaque root registry,
  immutable SHA-256 dataset CAS (six offline adapters), canonical StrategySpec,
  package-owned corpus snapshot + 14 current-fork scaffolds, import-free
  AST/security validator, hash-bound apply/run approvals, fixed-profile
  child-process runner, append-only session hash chain with recovery, independence
  audit, create-only native adapter installer.
• Native host adapters: claude-code, codex, opencode, openclaw.
• References, dev scripts, typed test suite + acceptance matrix.

Hardening pass (the four gaps closed)

1. Manifest freshness. manifest.json was stale (missing LICENSE/.gitignore;
   72 vs 74 files) and the package resources/distribution-manifest.json drifted on
   every source edit, failing the independence audit. Added scripts/build_manifest.py
   as the single regeneration entrypoint (package manifest written before root, since
   the root hashes the package manifest; verified idempotent) and a CI gate that
   committed manifests must match a fresh build.
2. Engine-root resolution. All 14 end-to-end cells failed because
   BACKTRADER_AGENT_ACCEPTANCE_ENGINE_ROOT defaulted to the repo grandparent (rarely
   a valid Backtrader source root). New resolve_acceptance_engine_root() checks env
   var → sibling backtrader/back_trader checkouts → installed backtrader
   package, with actionable guidance when none is found. Version assertion changed
   from hardcoded 1.3.0 to the descriptor's actual version, making the suite
   portable across Backtrader builds.
3. Listing commands + doctor engine detection. Added data list, session list,
   runs list, engine --list; doctor now reports registered engine roots and a
   hint when none is registered.
4. Docs. Documented renderer scope (P0 renderer is a scaffold selector; entry/
   exit/sizing/risk are validated and hashed but not translated to executable
   logic), plus engine-root auto-resolution. Added CI, examples/, SECURITY.md,
   CONTRIBUTING.md, CHANGELOG.md; .gitignore now ignores .backtrader-agent/.

Verification (run locally on this branch)

Gate	Result
pytest tests	59 passed
scripts/audit_independence.py	passed (6/6 checks)
scripts/doctor.py	ready
scripts/run_acceptance.py	passed: 14 cells × 2 modes, 6 adapters, crash/resume + repair gates, clean-wheel execution, sibling-absent enforcement
manifest freshness	committed manifests match build_manifest.py output

Test plan

• Local: pytest tests, audit_independence, doctor, run_acceptance all green
• Manifest freshness: second build_manifest.py run produces no diff
• Smoke: data list / session list / runs list / engine --list / doctor
  return correct results on empty and populated state
• CI green on Python 3.9 / 3.11 / 3.12 (pending first CI run)
• Reviewer confirms examples/ walkthrough is followable end-to-end
• Reviewer confirms no sibling product dependency (backtrader_mcp /
  backtrader_skills absent in clean install)

Notes

• The project baseline was never in git, so baseline and hardening are
  indistinguishable in the tree; organized into 3 area-based conventional commits
  (feat runtime, chore CI/governance, docs README/report).
• No live broker, no network data, no OS-sandbox claim — see SECURITY.md and the
  "Honest P0 limits" in README.