# Contributor lessons proposed by this build, not canonical instructions

## Decimal-to-rational checking must avoid context-rounded Decimal arithmetic
What happened: a stored factorial-normalization checker used Decimal.scaleb
before int(), silently reducing a 100-digit value to the default 28-digit
Decimal context. It reported false disagreements on correct coefficients.
What the skill says now: exact arithmetic must be exact, and checkers need
known-answer and negative controls; this specific parsing trap is not described.
What it should say: when turning decimal strings into exact rational endpoints,
extract Decimal.as_tuple() sign, integer digits and exponent without Decimal
arithmetic, or use another exact parser. Test a difference below 1e-100 so
the checker cannot hide a context-rounding bug behind ordinary short controls.
Evidence: stored_identity_check-failed-decimal-context.py,
offline-identities-failed-decimal-context.json, offline-identities.log,
stored_identity_check.py and stored-identity-results.json for T753. The
corrected 75 comparisons all pass; generator values were unchanged.

## Inspect slotted verification reports through their documented attributes
What happened: verify(sample=None) returned 42/42 matched with no differences,
but the run's subsequent vars(report) logging failed because VerifyReport uses
__slots__ and has no __dict__.
What the skill says now: use verify(sample=None), but does not describe the
report object's representation.
What it should say: use report.checked, matched, differing, missing, extra and
ok rather than assuming the report is a dict or has __dict__. Preserve a
successful numerical report when a surrounding logging adapter fails.
Evidence: verify-stored.log, clients/python/numberdb/_generate.py VerifyReport,
verify-stored-results.json, and corrected verify_stored.py for T753.

Existing proposal note: the older lesson file already discusses absent
ComplexBall.is_finite(). This run confirms that its real and imaginary parts
must be checked separately; it does not propose a duplicate lesson.

Artifact-run policy keeps this proposal here instead of editing the shared
agents/lessons tree or the canonical skill. A maintainer may consider it with
an appropriate regression test.
