Fix add_pvalue() matching Gray-test p-value to plotted outcome by name (#277)

[ddsjoberg]

What

Fixes #277 — for competing-risks fits, add_pvalue() could report the Gray-test p-value for the wrong competing event.

Why

In update_add_pvalue() (R/add_pvalue.R), the outcome position was derived from the row order of tidycmprsk::tidy(survfit)$outcome and then reused to index the wide p.value_* columns produced by tidycmprsk::glance(). Those two orderings are independent (tidy orders by first appearance of the label; glance numbers columns by failcode), so when they diverged the reported p-value belonged to a different outcome than the one plotted.

Reproduced by re-leveling death_cr so the factor order differs from the failcode order: choosing "death from cancer" (tidy position 1) wrongly pulled p.value_1 (0.766, belonging to "death other causes") instead of the correct p.value_2 (0.159).

How

The chosen outcome is now matched by name against glance()'s own outcome_* columns, and the paired p.value_* column is pulled. This removes any dependency on tidy() ordering. The now-unused tidycmprsk::tidy() call and position calculation in the single-outcome-check block were removed; the multi-outcome abort is unchanged.

Tests

• Added a regression test in tests/testthat/test-add_pvalue.R using a fit where tidy() and glance() orderings differ. It asserts each plotted outcome reports its own p-value, that the two outcomes' values are not swapped, and includes a sanity check that the orderings genuinely differ (so the test actually exercises the bug).
• Existing competing-risks tests only checked that some p-value string was present, so the bug previously passed them.

Reviewer notes

• The fix assumes the plotted outcome label matches a glance() outcome_* value exactly; the existing single-outcome guard and ggcuminc() outcome validation keep the no-match path unreachable in normal use.
• NEWS.md updated with a bullet under the development version.

🤖 Generated with Claude Code

[ddsjoberg]

Thanks for report @zhaohongxin0 ! Can you take a look at this PR, please? I think it should address the issue

[zhaohongxin0]

Thank you for the quick fix. I tested the PR head (ggsurvfit 1.2.0.9001) in an isolated R library using both the reproducible example from #277 and an independent competing-risk dataset.

The plotted outcome now receives the correct Gray-test p-value:

• Early recurrence: expected 0.496117, displayed 0.496117
• Late recurrence: expected 0.744450, displayed 0.744450

Both values exactly matched the corresponding outcome-specific values from tidycmprsk::glance(). The current CRAN release (ggsurvfit 1.2.0) selected the competing event's p-value in the same analyses.

The fix works as intended in my external validation. Thank you again.