docs(terminal): correct the tmux env-freshness rationale to one server per drive + sweep dead test

[roeikosh]

Description

Follow-up to #358 (review of the merged branch).

#358 removed new-session -e — env values on the tmux command line are
world-readable via /proc/<pid>/cmdline. Freshness now comes from starting a
session-private tmux server with that drive's scrubbed env, so the env travels
in the server's own environ. Seven sites still gave the deleted -e
mechanism as the reason a drive sees a rotated secret:

• docs/agent-tools/terminal-driver.mdx — the unsafeDisableSandbox
  paragraph, i.e. the text an operator reads to decide whether running without
  the jail is acceptable. It described a mechanism that no longer exists and
  that leaked env on argv.
• terminal-worker-backend-attach.ts — the load-bearing security comment
  directly above loadTmux.spawn, plus the module header and the
  unsandboxed field doc.
• terminal-spawn-plan.ts — all three sandbox-off sites.
• setup-terminal-tools.ts — the create-frame comment for the operator opt-out.

Also corrects two field docs #358 invalidated: tmuxSocket (absent now means
the worker derives tmux-<sessionId>.sock, not "the legacy single-socket
default" — a wrong pointer when debugging a failed re-attach) and socketPath
(documented one stable tmux.sock; production always passes the per-session
path).

Deletes the skipped SUPERSEDED: -e test: its body still passes an env field
that no longer exists on buildTmuxSpawnArgv's parameter type, an
excess-property error hidden only because packages/skills/tsconfig.json
excludes test files. It can never be un-skipped, so it is rot.

No behavior change.

Related Issue

N/A: documentation, comments, and one dead skipped test — no runtime behavior or
security expectation changes.

Type of Change

• Bug fix
• New feature
• Breaking change
• Documentation
• Refactor

Checklist

• Targeted checks for the changed area pass
• Full repository validation passes (pnpm validate) — see Additional Notes
• New or changed behavior has a test that demonstrated the RED state before the production change
  (N/A — no behavior change)
• Documentation updated (if applicable)
• No secrets or credentials committed
• Security implications considered
• The change is focused on one concern and links the related issue, or states the allowed N/A
  reason

RED Test Proof

EXEMPT: comments, docs, and removal of one already-skipped test. No
packages//src/* behavior change.

Evidence and Residual Risk

• Contribution path: not applicable (documentation of an existing security
  boundary; no code path changes)
• Evidence level: code trace — each corrected comment was checked against
  the post-fix(terminal): keep drive secrets off the tmux command line via one server per session #358 call path (runOneShot(a.env) carries the env; buildTmuxSpawnArgv
  no longer accepts one)
• Tested profile: Linux (Ubuntu 6.8), Node 22, no isolation change; docs:check,
  tsc -b packages/skills packages/daemon, terminal-driver unit tier (917 passed)
• Residual risk: prose accuracy is not machine-checked — nothing prevents the
  same rationale drifting again after a future change to the tmux backend. This PR
  does not add such a check.

Screenshots

N/A

Additional Notes

pnpm validate on my dev box: docs:check, build:clean, cycles (0 circular
across 16 dist packages), cycles:refs, and lint:security (0 errors) all pass.
test:coverage ends 60 failed / 41114 passed across 15 files, all in the
real-bwrap / live-PTY / jsdom tier that is broken on this host
(bwrap: loopback: Failed RTM_NEWADDR: Operation not permitted;
localStorage.removeItem is not a function in packages/web). This PR touches
no web code and changes no runtime behavior, so those cannot originate here —
but the per-package coverage floors never evaluated (the run aborted first), so
CI is the first place that gate actually runs.