…ports

Both variants run the same tasks, so an unpaired Welch test carries
between-task difficulty variance in its denominator and is systematically
underpowered. Add `paired_t_test` (one-sample t on per-task differences,
reusing the Phase 1 exact t CDF) and render a `## Paired Comparison`
markdown section for any 2-variant experiment.

The existing paired score comparison moves out of `## Replicate
Statistics` into that section and loses its `replicate_count > 1` gate,
so single-replicate A/B experiments — the common case — get it too.

Pairing is over per-task mean scores: replicate slots within a task share
the task effect and are not independent, so pairing them individually
would understate the standard error and manufacture significance (the
very failure this change set exists to fix). The task is the unit of
analysis, which is also what the section's note reports.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

- Guard the t-test helpers against non-finite input (rubric item 15).
  student_t_two_tailed_p fails closed to p=1.0; welch_t_test and
  paired_t_test return None, which renders as "—" rather than a
  fabricated p-value.
- Drop the "more powerful than the pooled Welch test" claim from the
  Paired Comparison note. Pairing loses degrees of freedom, so it is not
  unconditionally sharper — the 2-variant snapshot is itself a
  counter-example (Welch p=0.658 vs paired p=0.766). The note is now
  descriptive: pairing cancels between-task difficulty.
- Stop excluding tasks whose two variants ran unequal replicate counts.
  That constraint existed to zip replicate slots; pairing per-task means
  has no such requirement, so the exclusion only discarded real data
  (and biased the comparison toward fully-completed tasks).
- A 2-variant experiment with exactly one common task now says why it has
  no paired result instead of silently rendering nothing.
- Note in-code that the bootstrap CI and the exact-t p-value are separate
  inference models and can disagree at small task counts.
- Warn when the incomplete-beta continued fraction fails to converge
  rather than returning a partial value silently.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Neither is a ~30-minute guard: the non-finite-input rule needs dataflow
analysis rather than the syntactic matching the CExxx rules do, and the
bootstrap percentile indexing is latent pre-existing code with no caller
that can currently reach it.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Out-of-domain values previously surfaced as a bare IndexError from the
percentile lookup. Refuse them explicitly instead: clamping into range
would quietly return an interval of the wrong width, which is worse than
failing. No caller passes either parameter today, so nothing changes for
existing report paths.

Closes the harness candidate deferred earlier this run.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Three related report-correctness fixes:

- De-duplicate _collect_variant_series. The markdown and HTML reporters
  each had a copy and they had drifted: the markdown one divides
  duration_seconds by replicate_count (correct — VariantResult sums it
  across replicates) while the HTML one used it raw, so HTML showed
  replicate-inflated durations. Both now call one collect_variant_series
  in reports_stats.

- Replace the paired percentile bootstrap with a Student-t interval.
  The section printed a bootstrap CI beside an exact-t p-value — two
  inference models on the same quantity, which could disagree, and the
  bootstrap was degenerate over the 2-3 task means it now resamples. The
  CI derives from the same distribution as the p-value, so the two always
  agree about whether 0 is excluded. New student_t_critical (bisection on
  the existing t CDF) and paired_t_ci; paired_bootstrap_diff_ci is gone.

- Render the paired section in the HTML report too. Both reporters now
  render one shared reports_stats.paired_comparison, so they cannot
  disagree about the numbers.

Also adds the non-finite harness deferred earlier: it enumerates the
module rather than listing helpers by hand, so a helper added later is
covered automatically. It immediately caught regularized_incomplete_beta
returning NaN for NaN input, now an explicit ValueError.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>