fix(samples): de-overfit no_implicit_null_initial (inflector false positive)

[aspala]

Problem

PR #30's action comment flagged lib/codeqa/metrics/file/inflector.ex:32-45 — the detect_casing/1 function, a pure cond-dispatch with regex guards — as a CRITICAL no_implicit_null_initial block. There is no nil-then-assign anywhere in it.

Unlike the file-invariant phantoms fixed in #31 (every block of a file getting the same delta), this was a genuine per-block outlier — the relative-delta floor from #31 correctly let it through because it does stand out. The real cause is upstream.

Root cause

Same overfitting as #31: the no_implicit_null_initial behavior had one good/bad sample pair but a ~105-dimension scalar vector. Learned from a single data point, the vector encodes the quirks of that one file and matches dense cond/case blocks by chance.

Fix

Added 3 good/bad Elixir sample pairs (pricing, session, router), each contrasting:

• bad: x = nil then if … do x = … end across branches
• good: direct assignment via case / with / function results

Relearned the scalar vector via sample_report --apply-scalars (457 written, 148 deadzoned). No engine code changed — consistent with the de-overfit approach.

Result

• inflector.ex no_implicit_null_initial finding: 0
• Sanity-checked: the behavior still separates cleanly — its bad/loader.ex sample anti-aligns, good/loader.ex does not. No real detections lost (this codebase has no genuine nil-init-then-assign, which is idiomatic-Elixir-untypical).

884 tests, credo --strict clean.

Stacked on #31 (base fix/sample-overfitting-false-positives).

[aspala]

Superseded by #35, which re-lands the #31+#32 content directly on main (the stacked base chore/num42-refactors-scout was squash-merged via #30, orphaning these). All changes are now in main.