fix(#712): skip session-start-hook.sh's stdout buffer under an active trace

[fdaviddpt]

Round-2 windows-latest leg of #690. bash -x scripts/session-start-hook.sh was observed exiting 0 with zero stdout bytes on windows-latest in test_the_real_hook_traces_past_the_bootstrap, while six neighbouring bootstrap-dirs.sh-only tests in the same file pass on the same runner. No Windows runner was available to this lane (or to the original issue filer) to step through the actual Git-Bash fd mechanism directly, so the fix below is REASONED, not OBSERVED, on the one platform it targets.

session-start-hook.sh buffers its own stdout into a temp file (exec 3>&1; exec > "$_REMEMBER_CTX_FILE") so a plugin-promo systemMessage JSON can wrap the output, then restores the real fd and cats the buffer back at the end. This adds a guard that skips the whole buffer/fd-swap whenever an active xtrace (case "$-" in (*x*))) or REMEMBER_TRACE=1 is set -- mirroring the existing, already-shipped precedent in bootstrap-dirs.sh, which already stands its own fd 2 redirect down under the identical two conditions (don't shadow output an operator deliberately went looking for). When the guard fires, _REMEMBER_CTX_OK stays empty and the hook falls back to the exact code path already proven safe by test_unwritable_ctx_buffer_does_not_burn_the_cooldown (buffer unwritable -> print live, no promo, no cooldown marker burned). The only visible cost of the guard firing is that the plugin-promo banner does not show under an active trace; the memory context itself is never affected.

Also updates the stale xfail reason text on test_the_real_hook_traces_past_the_bootstrap to describe the fix and point at the new test class, while leaving the xfail itself in place (strict=False, scoped to sys.platform == "win32", raises=AssertionError) since there is no Windows runner available in this environment to confirm it flips to XPASS -- that confirmation has to come from the next windows-latest CI run.

New tests

tests/test_trace_not_swallowed_690.py::TestCtxBufferSkippedUnderTrace, using the plugin-promo systemMessage JSON as a cross-platform-observable proxy for whether the buffer engaged:

• test_positive_control_promo_shows_without_trace -- must-fire control: untraced, buffer writable, promo-eligible fixture -> systemMessage present.
• test_buffer_skipped_under_bash_x -- the actual guard: bash -x active -> systemMessage absent.
• test_buffer_skipped_under_remember_trace -- REMEMBER_TRACE=1 -> systemMessage absent, the same explicit opt-out bootstrap-dirs.sh already honours.

Testing

RED (scripts/session-start-hook.sh reverted, new tests kept): pytest tests/test_trace_not_swallowed_690.py -k TestCtxBufferSkippedUnderTrace -q --no-cov -> 2 failed, 1 passed (positive control passed; both guard tests failed against the old code).

GREEN: pytest tests/test_trace_not_swallowed_690.py -q --no-cov -> 10 passed, 1 skipped (pre-existing, unrelated BASH_XTRACEFD-version skip). pytest tests/test_plugin_promo_574.py tests/test_session_start_fork_tax_665.py tests/test_session_start_spawn_reduction_679.py -q --no-cov -> 27 passed, no regressions in neighbouring session-start-hook.sh suites. Full pytest (the whole test_command) was not run locally, narrowed per policy to touched files plus the guard-derived set.

Closes #712

[AI-generated]

[fdaviddpt]

Follow-up commit 22833e7 pushed after CI went red on windows-latest (job 105918977745).

The failing leg was this PR's own new positive control (test_positive_control_promo_shows_without_trace), which failed with assert "systemMessage" in '' and the job's own stderr showing session-start-hook.sh: line 213: ./resolve-paths.sh: No such file or directory.

Root cause: the test invoked the real hook as bash [..., str(SESSION_START)]. On Windows, plain str() on a WindowsPath renders a pure-backslash path with no forward slash at all. session-start-hook.sh's self-location logic (_HOOK_DIR="${BASH_SOURCE[0]%/*}") has no fallback for that shape, so with zero / in BASH_SOURCE[0] the strip is a no-op and _HOOK_DIR becomes . -- source "$_HOOK_DIR/resolve-paths.sh" then fails from whatever directory pytest happens to be running in, and resolve-paths.sh's own || exit 0 swallows that failure silently. This is very likely also the exact original #712 symptom (bash -x session-start-hook.sh exiting 0 with zero stdout bytes) -- a test/harness path-normalization bug, not a production bug (hooks.json always invokes via a literal path with at least one forward slash regardless of platform).

This is a known bug class in this repo: tests/test_windows_native_hook_cwd_448.py documents hitting and fixing the identical thing, and tests/test_hooks_json.py establishes .as_posix() as the standing convention. All four bash [..., str(SESSION_START)] call sites in tests/test_trace_not_swallowed_690.py now use .as_posix() instead.

The original fd-swap/trace guard from 16598ee is kept as a harmless, precedent-matching improvement (mirrors bootstrap-dirs.sh's existing convention), but the path-normalization fix is now believed to be the actual root cause of the CI symptom. changelog.d/712.fixed.md updated to describe both. Not independently reproduced on a Windows runner -- REASONED, confirmed by the next windows-latest CI run rather than locally.

Note: this branch's own CI may still show unrelated red from #734 (a stale _SANCTIONED_DIVERGENCE entry on main at 7ba7609, tracked separately) until that lands -- unrelated to this fix.

[AI-generated]