👷 ci: hold native coverage to LLVM's total

[gaborbernat]

The native coverage gate reads a narrower thing than its name suggests. --fail-uncovered-lines 0 counts a line that no function reached, and cargo-llvm-cov drops a line from that list as soon as any function or macro expansion on it ran, which cargo-llvm-cov#404 describes. A physical line shared by a covered call and an uncovered closure leaves the list while LLVM still counts it missed.

A ten-line crate shows the size of that gap. One function, v.map_or(0, |x| x * 2), tested only with None, puts the file at six covered lines of seven. cargo-llvm-cov reports 85.71%, prints no Uncovered Lines section at all, and the gate exits zero. 📉 Adding --fail-under-lines 100 to the same profile exits one. On this workspace the two measures agreed everywhere I sampled, so the divergence is rare rather than absent, which is what makes it worth a gate rather than a review habit.

The gate also now reports to the terminal instead of into the lcov file. In lcov mode a failure prints nothing whatsoever, so a red gate would hand back an exit code and no file to look at. Running the report twice over one profile costs a second report step and none of the measurement, which is where the sixteen minutes go. A failing file then shows in the per-file table, and a file carrying missed lines while listing none under Uncovered Lines is the shared-line kind. That comparison is the only way to locate them, since LLVM does not expose those source lines and llvm-project#126307 tracks the omission.

The contract that already checks this recipe's target directory now also holds both properties. Removing the flag fails it, and folding the gate back into the lcov write fails it, so the gate cannot quietly return to reporting nothing.

Two things for a reviewer. This PR's own coverage job is the measurement that matters: it runs just coverage-native on x86_64-linux, which is the only platform this gate covers, and it will say whether the workspace already sits at LLVM's 100% or whether files remain. If it comes back red, the per-file table now names them and the remediation belongs in its own change rather than buried here. I did not run the full native profile locally, since it is the machine-wide heavy command the brief reserves.

Closes #1745

[codspeed-hq]

Merging this PR will not alter performance

✅ 29 untouched benchmarks
⏩ 133 skipped benchmarks¹

---

_{Comparing fix/enforce-total-line-coverage-1745 (227835c) with main (892226e)}

Footnotes

1. 133 benchmarks were skipped, so the baseline results were used instead. If they were deleted from the codebase, click here and archive them to remove them from the performance reports. ↩