[DanielNetzer]

Implements issue #290.

Molecule: mol-290-d0cb51462659714fd4b190eb
Formula: forgecell-github-issue
Base: 777b56eb63c6fd85854fec73a405fb9c545a8e5d
Delivery: 2f59dfa9ea6247e94f3bbdadf6b4d61e5cbf3358

Awaiting exact-commit CI and human review. Harness-reported checks are not independent CI evidence. No merge, deployment or issue closure is authorized.

[DanielNetzer]

Adversarial review (plugins slice 2a): no HIGH findings and no correctness MEDIUMs. I checked the framing bounds, the cap on skipped messages, id matching, refusal of server-initiated requests, duplicate and case-variant key rejection, the contract check on initialize, the strict decoders, Error() (fixed reasons only, server text redacted), and that the safetext move matches the base sanitizer and redaction behaviour.

Two hardening MEDIUMs and four LOWs are tracked in #294, which lands before slice 2b:

• the process group isn't killed after a clean exit;
• cleanup of a chmod-ed run directory fails silently;
• the interpreter isn't re-verified just before exec;
• Initialize doesn't refuse a mismatched protocolVersion, doesn't sanitize the reported contract, and has no default timeout;
• U+2028 and U+2029 pass the sanitizer and identifier checks;
• the conformance bound checks pass when no tool was called.

Another LOW: TestPluginTestAgainstTheFakeServer takes about 48s, and about 13s even on an idle machine. Building the fake server once per package with TestMain and sync.Once would make it faster.

Integration: I applied this branch to current main (with #291 and #295) and go vet ./... passes. These packages also pass: go test -timeout 20m ./internal/{plugin,extension,cli,safetext,ui,forgesync,delivery,forge}/....