fix(tests): keep Dependencies/.env out of the master suite (and fix the loader guard)

[DoRmAmMu1997]

A guard from #181 fails on your machine — and only there

Running the gates on merged main in your checkout: 617 run, 1 failure — test_the_flag_is_set_for_the_load_only, one of the guards #181 added. It found SL_HUNTING_ENABLED='true' where it had been unset before the master loaded.

The loader isn't the cause. It sets the flag inside patch.dict, which restores the environment when the load ends. The value comes back from the next import in the test module: flattrade_execution.py calls load_dotenv() at import time, outside any patch, and re-reads Dependencies/.env. Proven in isolation — importing that one module moves SL_HUNTING_ENABLED from unset to 'true'.

CI and every worktree have no .env, which is why the guard passed on #181's CI.

The fix

Record the value the moment the master's load block ends — before that import — and have the guard (renamed test_the_loader_puts_back_the_flag_it_set) compare that to the pre-load value. It now tests exactly what it claims, and doesn't care about .env loading elsewhere.

Proof

Reproduced without copying any credentials — a throwaway Dependencies/.env containing only SL_HUNTING_ENABLED=true:

	old guard	new guard
with the one-line .env (your condition)	1 failure	617 OK
without a .env (CI's condition)	617 OK	617 OK

Negative-tested in both conditions: setting the flag for the whole process before the load, and setting it after the load instead of inside patch.dict, are both caught; a control rewording passes. The throwaway .env was removed.

Gates: master 617 OK with and without the one-line .env · 28 market-data-health · 1594 pytest · ruff · mypy (80 files) · compileall.

---

Second commit: keep Dependencies/.env out of the master suite entirely

The first commit worked around the symptom. This removes the cause: on your machine the master suite ran against your .env; CI, which has none, didn't.

Two leaks, not one. The master, dhan_execution and flattrade_execution all call load_dotenv() at import.

1. The Flattrade imports ran outside any patch, copying your whole .env — flags and broker credentials — into os.environ for every later test.
2. The master's load was already inside patch.dict, which restores os.environ — but its load_dotenv still ran, and still let .env values set the master's import-time constants.

Measured without touching your real .env — a throwaway .env holding the 616 key names from env.example, each set to a placeholder:

	leaked into os.environ	tests that changed
before (this PR's parent)	616	21
Flattrade imports wrapped only	0	12
all three loads isolated	0	0

Of the 21: nine read the environment at run time (broker routing ×4, the live/paper gates, the per-strategy toggle), twelve depended on master constants set at import (websocket feed, strangle re-entry, order correlation). None needs a .env — all pass without one — so none was relying on it; they were exposed to it. They passed on your box only because your values are sane.

The fix is test-side only. Every module-level repository load goes through _isolated_from_dotenv: patch.dict restores os.environ exactly (carrying the values a load needs), and dotenv.load_dotenv is stubbed so the file isn't read at all. Production's load_dotenv is unchanged; no default moves. Follows the precedent in test_index_fetch_construction.py.

Guards — one of which holds in CI, where the others would be blind:

• behavioural: os.environ snapshotted before the first repo import and after the last; fails naming any variable added/removed/changed — names only, never values, since on your machine a value can be a credential;
• probe: a throwaway module + .env in a temp dir, loaded with the real load_dotenv — bare, it must pick the value up (the control); isolated, it must not;
• structural: the master and both Flattrade loads must go through the isolation — so a bare exec_module fails here instead of passing CI and misbehaving on the trading box.

Negative-tested in both conditions (no .env, and the 616-key throwaway): reverting the master load, reverting the Flattrade load, dropping the stub, and dropping the restore — all four caught, each by the guard meant for it; with the throwaway .env the reverted fixes also bring back the original failures. Control passes. The throwaway .env was removed every time.

Gates (this commit): master 620 OK, identical with and without a .env · 28 market-data-health · 1594 pytest · coverage report exits 0 under the 73% floor · coverage policy passed · ruff · mypy · compileall · bandit.

🤖 Generated with Claude Code