test-coverage always fails: a truncated covr trace file aborts merge_coverage and skips the codecov upload

[mattfidler]

The test-coverage workflow has been failing on main and on essentially every open PR branch. It is not any one branch's fault, and the test suite itself passes -- the run ends [ FAIL 0 | WARN 31 | SKIP 58 | PASS 76505 ] and then dies during covr's merge:

Error in readRDS(f) : error reading from connection
Calls: <Anonymous> ... merge_coverage.character -> lapply -> FUN -> as.list
Execution halted

Because the step exits non-zero, codecov/codecov-action is skipped, so no report is uploaded at all. Each failed run burns 2--5 hours first.

Cause

covr:::add_hooks() appends this to the installed package's load script:

reg.finalizer(ns, function(...) { covr:::save_trace(Sys.getenv("COVERAGE_DIR", "<lib>")) }, onexit = TRUE)

So every R process that loads the instrumented rxode2 writes a covr_trace_* file into the merge directory when it exits -- including the helper R subprocesses the test suite spawns (test-omp-num-threads.R, test-pkg-exported-funs.R, ...). package_coverage() then does:

trace_files <- list.files(path = install_path, pattern = "^covr_trace_[^/]+$", full.names = TRUE)
coverage <- merge_coverage(trace_files)

and merge_coverage.character() is just lapply(x, function(f) as.list(readRDS(f))) -- no error handling. One partial file takes down the entire run.

Evidence

Downloading the coverage-test-failures artifact from a PR run (31139091448) and a main run (31040944854) gives the same picture both times: 25 trace files, exactly 8 unreadable. The bad ones have a valid gzip header and real compressed data that simply stops (gzip: unexpected end of file), at a size that is an exact multiple of 4096 -- a write that was cut off, not a corrupt file.

Fingerprinting the readable traces shows the truncated ones cost almost nothing:

trace	covered expressions
1 file (the testthat run, 26 MB)	15972 -- intact
6 files	0
4 files	41 (.onAttach only)
6 files	~525 (rxForget)
8 files	truncated

All the real coverage is in the one intact trace. The 24 others are near-empty helper processes.

Fix

Confirmed against the real artifact files: unpatched, merge_coverage() errors; with unreadable traces skipped it returns a normal coverage object with 16027 of 20218 expressions covered.

PR #1208 makes the workflow drop unreadable traces and report which ones, so a poisoned helper trace costs a little coverage instead of the whole report.

That is deliberately a tolerance fix, not a root-cause fix -- it does not explain why those 8 subprocesses die mid-write. Worth a separate look, but the job should not be red in the meantime.