ci: capture compile stderr in parity + compile-smoke jobs

[TheHypnoo]

Why

11 tests are tracked in test-parity/known_failures.json with status: \"ci-env\" and the same vague reason — "macOS-14 CI runner SDK/linker version interaction":

• test_gap_buffer_ops, test_buffer_small_alloc, test_buffer_numeric_read_intrinsic, test_gap_node_crypto_buffer, test_gap_typed_arrays, test_stress_buffer, test_inline_uint8array_param, test_issue_167_loop_alloca_stack_eat, test_issue_227_array_buffer_bytes, test_gap_fetch_response, plus the new test_issue_234_blob_methods from fix(fetch): #234 real Blob with arrayBuffer/text/bytes/slice instance methods #238.

But the actual error message has never been pinned down, because both compile harnesses silently discard stderr:

• run_parity_tests.sh:218 captures compile_output into a local var but only logs \"compile error\" with no detail.
• .github/workflows/test.yml::compile-smoke compiles with 2>/dev/null, so the job summary lists only the failing test names.

Net effect: every PR that adds a Buffer / Uint8Array regression test ends up in the same skip list ("ci-env / passes-locally") on inference alone, and the underlying root cause never gets fixed. The v0.5.354 changelog itself acknowledges:

"the underlying SDK/linker gap is the real long-tail problem and a CI runner image bump (macOS-15) would clear all 9+ ci-env Buffer skips at once."

…but that hypothesis has never been verified against an actual error message.

What this PR does

Pure diagnostic plumbing — no source changes, no SKIP_TESTS edits, no runner bump, no version bump.

1. run_parity_tests.sh

On compile-fail, write the captured compile_output to test-parity/output/<test_name>.compile_error.log. The existing per-test loop is unchanged; this just persists the stderr the script was already capturing into a memory-only var.

2. .github/workflows/test.yml

Two new artifact upload steps (both if: always() + if-no-files-found: ignore so green CI runs produce no artifact noise):

• parity-compile-errors-${{ runner.os }} from the parity job — gathers the per-test logs the script now writes.
• compile-smoke-error-logs from compile-smoke — captures stderr per test to /tmp/perry_smoke_logs/, plus inlines head -n 30 of each failure log into the job output (so a quick scan in the GitHub UI reveals the error without needing to download the artifact).

What you'll see on the next CI run

The parity job naturally exercises every ci-env test (it doesn't have a SKIP_TESTS list — it tracks them via known_failures.json instead). With this PR, that job will produce ~11 *.compile_error.log files in the parity-compile-errors-macOS artifact, one per known failure. Reading any of them gives the actual link error.

Decision tree once we have the data

• If the error matches the maintainer's hypothesis (link symbol mismatch / SDK ABI skew): file a follow-up PR bumping runs-on: macos-14 → macos-15 and removing the skips.
• If the error is a real Perry bug (e.g. a stdlib codegen path that emits something the macOS-14 SDK linker rejects but local does accept): fix it in Perry rather than skipping it.
• Either way: the diagnostic capture stays as defensive infra — future Buffer/typed-array regressions will surface the link error in the PR's CI run instead of getting silently added to the skip list.

Refs

• v0.5.354 changelog ("the underlying SDK/linker gap is the real long-tail problem")
• v0.5.314 / SIGTRAP crash when a program mixes Buffer-typed and Uint8Array-typed function parameters #169 (test_inline_uint8array_param — original ci-env entry from this family)
• fix(fetch): copy body bytes into response.arrayBuffer() result #232 / response.arrayBuffer() returns metadata-only object — body bytes never reach TS #227 (most recent additions to the skip list)
• fix(fetch): #234 real Blob with arrayBuffer/text/bytes/slice instance methods #238 (PR for response.blob() drops body bytes — same shape as #227, needs Blob instance methods #234, latest addition to the same skip list)