[fix] preserve session recordings and report write failures (STAR-106)

[vt128]

--record omitted readline prompts/input echoes despite promising a complete session, leaked two pipe readers per completed recording, and treated a full recording destination as success. A runtime write failure could also stop draining a pipe and stall a sufficiently chatty script.

Tee readline's existing platform-specific terminal writers without taking ownership of them, close all owned pipe ends on normal and failed setup paths, and retain recording errors while continuing to drain script output. Return a non-zero exit code on recording failure while preserving an existing non-zero script exit code. Stop is idempotent. The readline version is unchanged; its existing dependency is now a direct import.

Fixes STAR-106. Behavior changes: transcripts now include terminal prompts/input echoes and may contain ANSI sequences; recording setup/write/close failures are observable. The real terminal rehearsal now runs for the installed Unix release archives, including plain REPL, post-script inspection, evaluation-error recovery, Ctrl-C and EOF. Portable session append/error tests run on all six targets.

Validation:

• Before: a real PTY omitted prompts; 20 recordings increased open descriptors from 7 to 47; /dev/full returned exit 0 and executed the script. Regression tests failed before the fix.
• After: full build/race twice/vet/format; Go 1.26.8 Docker build/race; actual runtime Dockerfile build/run and disk-full failure; real macOS PTY REPL and inspection passed.
• Unit tests cover header errors/short writes, both pipe allocation failures, stdout/stderr forwarding failures, write/short-write/close failures, repeated stop, all classified exit codes, resource ownership and continued draining beyond pipe capacity.
• Existing tests remain in the functional test files; release CI repeats native archive installation and rollback tests.

Additional validation: ten alternating 4 MiB recording workloads (4096 print calls) showed 22.68 ms before and 22.77 ms after at the median on macOS arm64, including process startup (about +0.4%, a local sample rather than a throughput guarantee). Native CI also exposed non-portable descriptor-test assumptions, which are now checked using the documented close contract. The terminal driver waits for readline raw mode before control keys and drains terminal output through child exit.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 87.99%. Comparing base (760b2d2) to head (6db0fe1).

Additional details and impacted files

@@            Coverage Diff             @@
##           master      #37      +/-   ##
==========================================
+ Coverage   86.34%   87.99%   +1.64%     
==========================================
  Files          23       23              
  Lines        1128     1166      +38     
==========================================
+ Hits          974     1026      +52     
+ Misses        117      108       -9     
+ Partials       37       32       -5     

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.