Port Bun shell environment propagation semantics

Follow-up from #689 and the exact Bun shell fixture at checkout 4982b91.

Remaining unchanged-suite failures:

• bunshell > variables > cmd_local_var
• bunshell > variables > shell var
• bunshell > variables > export var
• bunshell > variables > syntax edgecase
• deno_task > change env

The observed child environment diverges from Bun semantics, including PWD propagation and FORCE_COLOR handling. Port the upstream environment model and assignment/export behavior logically; do not change tests, skips, deadlines, workloads, or assertions.

Strict baseline after e83b67b: 369 pass, 83 upstream TODO, 9 fail across 461 tests. Related trackers: #66 and #202.

Resolved on main by 0bc37d2. The shell environment maps and assignment/export behavior were already correct; the native corpus runner changed into packages/runtime/test without synchronizing PWD, so the unchanged Bun fixture constructed a stale expected environment. The runner now updates and restores PWD together with cwd. All five affected unchanged upstream tests now pass: cmd_local_var, shell var, export var, syntax edgecase, and deno_task change env. Full fixture result improved from 369 pass / 83 todo / 9 fail to 377 pass / 83 todo / 1 fail.