fix(compliance): review follow-ups to #270 — re-pin the admission floor, stop a NaN crash

[eaitbrahim]

Follow-ups from the independent adversarial review of #270. That PR was merged before these landed, so they come as a separate change against current main.

1. HIGH — the admission floor lost its only value pin

#270 changed the floor assertion from DEFAULT_MIN_QUOTE_24H_VOLUME == ScreenPolicy().min_median_daily_volume to <. The relationship guard is right and stays. But it was also the only test pinning the admission floor's value, and nothing replaced it.

Verified on main before this change: editing ScreenPolicy.min_median_daily_volume from 1000000 to 200000 — a 5× loosening of the criterion deciding which assets a money-moving tool may buy — passed all 2756 tests.

This adds test_the_admission_liquidity_floor_is_pinned_to_its_actual_value alongside the < test. The two are not redundant: one guards the relationship (discovery must never be stricter than the gate), the other guards the criterion itself.

Mutation proof on this branch:

floor -> 200000 :  1 failed, 2760 passed
  FAILED test_the_admission_liquidity_floor_is_pinned_to_its_actual_value
reverted        :  2761 passed, 1 skipped

2. MEDIUM — NaN volume crashed assets discover outright

The parse was inside a try, the comparison was not. Decimal("NaN") parses fine, then < raises decimal.InvalidOperation and takes the whole command down. Separately "Infinity" parsed and silently became a candidate.

Both are now counted as unreadable_volume, with the reasoning commented. Verified end to end:

survived crash -> ['GOOD']
unreadable counted -> 2
accounting: 1 + 2 == 3 -> True

3. MEDIUM — the survivor-count identity was load-bearing but unguarded

render_discover_report derives survivor_count = venue_product_count - excluded.total. Nothing asserted the underlying invariant, and an inconsistent DiscoverReport rendered nonsense (10 venue products -> -89 candidates). Adds a test pinning len(candidates) + excluded.total == len(products), and makes the subtraction defensive.

4. LOW/MEDIUM — the incident record was wrong, and the regression test pinned the wrong end

The comments described the 852,133–979,000 cluster as including ALGO. Measured against the live venue, the cluster is ATOM, AAVE, BCH, CRV; ALGO was a separate low outlier at 437,712.

Worse, the regression test pinned only 852133 — the top. A future floor of 500,000 would have passed it while silently re-hiding ALGO. Re-pinned to ALGO's 437712, which is what actually constrains the floor. Verified: a floor of 500,000 fails the corrected test and would not have failed the old one.

5. LOW — frozen=True was cosmetic

DiscoveryResult and DiscoverReport were frozen=True but held a mutable list[Candidate] — appendable and unhashable. Now tuple[Candidate, ...].

Gates

ruff check keel tests packages   All checks passed!
mypy                             Success: no issues found in 224 source files
pytest -q                        2761 passed, 1 skipped

🤖 Generated with Claude Code