fix(printf): implement %n; report C-library conversion failures like bash

[brianjfox]

Summary

Fixes the two bi_printf mismatches from #665.

%n

Previously printed the literal spec and consumed no argument. Now, matching bash's printf.def case 'n':

• binds the named variable to the byte count written so far in the current format pass — bash zeroes its tw counter on every reuse of the format string, so printf 'ab%n' u v w x sets each variable to 2
• a missing/empty name is silently ignored (exit 0)
• an invalid identifier (including a[0]) reports printf: `NAME': not a valid identifier and stops immediately with status 1, still writing the accumulated output (printf 'ab%ncd' 0 prints ab)
• a failed binding (readonly variable) reports through Shell::set but leaves the exit status 0, as bash does; namerefs are followed
• printf -v out 'ab%n' u counts bytes in the -v buffer the same way

C-library conversion failures

When snprintf refuses a conversion (field width past INT_MAX is EOVERFLOW on glibc; an unallocatable buffer is ENOMEM), gnash kept processing, printed a hardcoded macOS-flavored message at the end, and threw all output away. bash's PF macro stops immediately, reports strerror(errno), and still writes what was accumulated — gnash now does the same. (On macOS libc the huge-width case never fails — printf '%200000000000000000000s' x prints x/exit 0 in both shells — so this path mainly aligns Linux behavior.)

Verification

Byte-identical stdout/stderr/exit vs reference bash 5.3.15 on 14 %n cases (mid-format stop, format reuse, -v interplay, nameref, readonly, array-element and other invalid names, %5n) plus width/precision/%q/%5.2f/%(fmt)T regression cases. printf.tests scoreboard improves 393 → 388 diff lines (the %n mismatches are gone; remaining diffs are pre-existing option-parsing, -v validation and %.*s star-width gaps). smoke.tests passes.

Closes #665

🤖 Generated with Claude Code