chore: architecture docs, tests consolidated under Tests/, coverage floor 54.7% → 68%

[DoRmAmMu1997]

Two structural changes plus a gate tightening. No runtime behaviour is touched ? no strategy, execution, market-data or risk code changed.

1. docs/ is now a real architecture set

The folder held only Superpowers session artefacts (a plan and a spec from one July session, already stale). Those are now gitignored and untracked ? kept on disk, matching the existing .superpowers/ rule that had simply missed this subtree.

In their place:

Path	Contents
docs/README.md	Index and reading order
docs/hld/system-overview.md	Requirements, context, component + thread model, data flow, failure model, trade-offs, what would force a redesign
docs/lld/	12 component designs
docs/adr/	11 decision records

The ADRs are retrospective and carry the part that is expensive to reconstruct later ? why an acknowledgement is never treated as a fill, why live trading needs two flags, why the size multiplier is deliberately not global, why the LLM agents own judgement but never sizing or stops. Each one states the options actually considered and what was given up.

The division of labour against existing docs: README.md stays operator setup, CLAUDE.md/AGENTS.md stay agent rules, per-folder Readmes stay folder-local. docs/ answers "how does this fit together, and why".

2. Every test now lives under Tests/, mirroring the source tree

36 test files and 4 conftests moved. Runtime folders now contain only runtime code ? Dependencies/ previously showed 13 test files against 15 runtime modules.

Tests/
??? test_nifty_multi_strategy_master.py     ??? Signal Generators/
??? test_market_data_health.py              ?   ??? CPR AI Agent/
??? Data Extractors/                        ?   ??? CPR Strategy/
??? Dependencies/                           ?   ??? Regime Adaptive Strategy/
    ??? Dhan API/                           ?   ??? SL Hunting AI Agent/
    ??? Flattrade API/                      ?   ??? Subhamoy Strategies/

Every __file__ anchor was rewritten to resolve from the repository root. Each mirrored folder for a spaced-name source folder carries a conftest.py that puts the source folder on sys.path ? never the test folder ? so tests exercise the same import resolution the runner uses.

That narrow scoping is deliberate and is why a flat Tests/ was rejected: Signal Generators/CPR AI Agent/conftest.py inserts only that agent's directory precisely so a test cannot resolve an import production would fail on. One flat conftest would have erased that property for every strategy at once.

Two rules for new tests, documented in the README and CLAUDE.md/AGENTS.md: mirror the source path, and keep the filename unique repository-wide (pytest keys modules by basename ? there are no __init__.py files).

3. Coverage floor raised 54.7% ? 68%

54.7% was the baseline measured before the gate was installed. Hosted CI now measures 69.1%, so the final floor is 68%: a 13.3-point improvement with about 1.1 points of headroom against the environment that actually enforces the gate.

A local machine with neo_api_client and pyotp installed runs seven optional tests that CI's core environment skips and reads about 71.4%. Those tests affect modules outside the broker-adapter threshold table, so passing the adapter's separate 80% budget does not make the repository-wide totals equivalent. The floor must therefore be judged from hosted CI, not from a developer machine with extra optional SDKs.

The floor is documented as one-way: it only ever moves up, and only after a hosted CI run shows headroom.

Verification

Counts were compared exactly, not just checked for green ? a silently uncollected test still looks like a passing build.

Suite	Before	After
unittest master	487	487
unittest market-data-health	26	26
pytest	1089	1089

Also green locally: ruff, mypy (53 files), compileall, bandit, pre-commit config validation, the branch-enabled coverage run at the new 68% floor, and the per-module policy gate. Verified separately that zero Tests/ files leak into the coverage report ? the omit patterns needed both Tests/* and Tests/**/* because Coverage.py's * does not cross a path separator.

CLAUDE.md and AGENTS.md remain byte-identical after the ## What this project is marker, as test_repository_policy requires.

Note before merging

Tests/test_sl_hunting_schema.py differs between main and the unmerged feat/sl-hunting-knowledge-v4d branch. That PR will hit a rename-vs-edit conflict on the file's new path ? straightforward to resolve, but worth sequencing deliberately.

Follow-up (not in this PR, to keep it surgical)

test_repository_policy.py already fails the build on stale worker-roster claims in README.md, CLAUDE.md, AGENTS.md and the master file. Extending that staleness gate to cover docs/hld/ is recorded as an open item in ADR-0011.

?? Generated with Claude Code

[DoRmAmMu1997]

Architecture review follow-up implemented in commit 90d19c0.

Changes made:

• corrected HLD requirement F5 so it no longer claims unattended restart recovery; partial/lost-response tracking is in-process, while restart exposure is blocked by the account-wide startup audit and requires operator reconciliation
• documented the distinct exposure scopes: startup is account-wide, shutdown is gated by the runner-owned execution ledger, and the final account-wide audit is advisory for possible manual positions
• replaced the inaccurate "exactly two shared mutable objects" claim with the actual synchronized boundaries: the shared store/safety aggregate, serialized broker access, lifecycle and shutdown signals, and the optional Telegram queue
• clarified that CPR RSI/SRSI and geometry values are code-owned invariants covered by focused tests, not env.example settings
• aligned the lifecycle module docstring and strategy-worker LLD with those contracts
• corrected the PR title/body from a 70% coverage floor to the final hosted-CI-based 68% floor and documented why local optional SDKs read higher

Verification:

• repository policy: 11 passed
• trading lifecycle: 6 passed
• combined policy/lifecycle rerun: 17 passed
• CPR AI suite after restacking fix(cpr-ai): align authenticated Codex turn contract #117: 68 passed
• Ruff, compileall, pre-commit, and git diff checks passed

These are documentation/docstring corrections only; no strategy, market-data, risk, or execution behavior changed. PR #117 has also been rebased onto this new #116 head.