Fix hush-file-write: pr to a file stream must not be silenced under -q

[pyrex41]

Under -q (*hush* true), pr to a file stream wrote a zero-byte file: the canonical kernel pr gates all output on *hush*. An explicit write to a file is user intent, not the stdout chatter *hush* exists to silence.

Adds a native pr override in prims.lua that silences only when *hush* is set and the target stream is *stoutput*; for any other stream it temporarily clears *hush*, delegates to the original compiled-KL pr (byte-identical write path), and restores *hush*. Canonical kernel sources untouched.

(Note: the earlier fix/hush-pr-file-22 was never actually merged to main; this re-applies the fix cleanly on the current line.)

Verification

• bifrost hush-file-write: green for shen-lua; full corpus 30/30 across 6 impls.
• make test: 451 pass, 0 fail. make certify: 100%.

🤖 Generated with Claude Code

[pyrex41]

Summary

The PR implements a targeted native override for the kernel pr (defined in klambda/writer.kl as an unconditional (if (value *hush*) S (write...)) gate) inside install_native_stdlib. The wrapper consults *hush* only for the exact *stoutput* stream object (using Lua table identity); for all other streams (including files opened via (open ... out) and used by write-to-file), it temporarily clears *hush*, delegates to the original compiled-KL pr (so the write-chars/write-byte path and return value are byte-identical), and restores state. Updated CLI subprocess tests in cli_spec.lua now assert both that file payloads are written under -q and that pr to (stoutput) remains silenced. The change aligns shen-lua with shen-cl/shen-go/ShenScript for file writes while preserving the documented quieting behavior for console output. The implementation is careful about state, call-time lookups, and APP/PARTIAL currying paths; the dominant remaining risk areas are untested *sterror* interactions and any future stream-redirection scenarios.

Issue counts by severity

• bugs: 0
• suggestions: 1
• nits: 3

[pyrex41]

[suggestion] The silencing condition uses a strict identity test st == GLOBALS["*stoutput*"]. This means pr writes to *sterror* (and any user-redirected or other streams) are no longer silenced under *hush* (the second if GLOBALS["*hush*"] branch forces a write). The kernel pr previously silenced all streams; the PR description and comment focus exclusively on "standard output" vs. files, but *sterror* behavior changed as a side-effect. No test exercises (pr ... (sterror)) (or equivalent) under -q.

Suggestion: Either extend the exemption (or the "do not hush" policy) to also cover *sterror* if errors/diagnostics should remain visible under quiet, or add an explicit or st == GLOBALS["*sterror*"] + comment if stderr should also be hushed. Add a cli_spec or io_spec case for stderr under -q to lock in the chosen policy. Update the long comment at 741-751 to explicitly state the *sterror* rule.

[pyrex41]

[nit] The force-write path for non-stdout streams under hush uses pcall + manual error(res, 0) re-raise after restore. While this correctly restores *hush* even on error (and matches patterns elsewhere e.g. boot.lua:703), the level-0 error call can alter the reported source location / traceback relative to what a direct call to the original orig_pr would have produced. In normal success paths this is invisible.

Suggestion: Consider a local function force_pr(...) local saved=...; GLOBALS["*hush*"]=false; local res = orig_pr(...); GLOBALS["*hush*"]=saved; return res end and pcall only the wrapper (or just let errors propagate if the restore can be done in an xpcall finally-style). Add a comment explaining why pcall is required here vs. other native overrides.

[pyrex41]

[nit] The new "other half" test constructs an expression that calls pr to (stoutput) but relies on -e result printing (the launcher io.write(shen.call("shen.app", ...)) at bin/shen:96) to produce observable stdout for the non-hushed case. The test correctly uses a distinct return value (99) so the marker can only come from the pr itself, and it captures combined stdout+stderr. However the local outn = run(...) discards the exit code (unlike the -q case) and the assertions only check marker presence/absence rather than also asserting that the result echo still appears.

Suggestion: Capture and assert the exit code for the non--q case too (for symmetry), and/or assert that the "99" result echo is present in outn (to document that -e value printing is intentionally outside the hush gate). Minor robustness improvement only.

[pyrex41]

[nit] The install("pr", pr, 2) call (and the preceding local function install...) was inserted before the install("variable?", ...) line that was previously first. While install order is irrelevant (pure F/FA table writes with no side effects or dependencies among these overrides), the diff moves a pre-existing line.

Suggestion: No functional change needed; purely cosmetic. If desired, keep the original install order and place the pr install at the end of the block for minimal diff.