fix(e2big): one spawn facade + copilot prompts on stdin end the recurring E2BIG (#2640)

[rysweet]

Problem

"Argument list too long" (E2BIG / os error 7) kept recurring across Simard's agent launch sites even after two piecemeal fixes (#2660 decision-cycle, #2700 journal) and resurfaced on the SIGNAL channel — because there was no single chokepoint every large-payload launch was forced through. The root pattern: a large prompt/context inlined into argv (or expanded via -p "$(cat FILE)") overflows ARG_MAX and execve fails with E2BIG before the child runs.

Fix: one spawn facade

New simard::spawn_payload — the single chokepoint that composes the two already-shipping transports under one policy (a payload ≥ ARGV_PAYLOAD_MAX_BYTES = 8 KiB never rides argv/envp):

• copilot prompts → stdin via attach_prompt_std/attach_prompt_tokio (forces PromptDelivery::Stdin; copilot reads its prompt from stdin when no -p is given, and its -- positional would be misparsed as a subcommand — so a prompt of any size stays off argv).
• recipe context → file path via recipe_context → -c <key>_path=<abs> (composes recipe_context_file::ContextFile).
• pre-exec spawn errors → record_spawn_failure → errno-keyed classify_spawn_failure into overseer::failure_sink (no silent swallow).

Copilot prompt sites routed through the facade (the real recurring failure)

Site	Before	After
base_type_copilot meeting turn (backs the SIGNAL channel)	apply_std(Stdin)	routed through attach_prompt_std
engineer_loop::run_engineer_subprocess (Copilot)	inlined -p <prompt> in argv, stdin(null)	prompt-less argv (copilot --subprocess-safe --allow-all-tools --allow-all-paths) + prompt streamed on stdin from a feeder thread

The #1717 permission contract (both flags in order + COPILOT_ALLOW_ALL=1) is preserved; the RustyClawd path is untouched (not production-wired).

Anti-regression guard

tests/e2big_argv_guard.rs fails CI if a $(cat argv-expansion, an inline unbounded recipe key (day_context/draft/episodes/pr_body), or removal of the spawn_payload facade is introduced.

Whole-repo launch-site audit

Tier	Site	Disposition
A	base_type_copilot meeting/signal	stdin via facade — ships
A	engineer_loop/agent_spawn.rs copilot	stdin via facade; prompt-less argv — ships
A	ooda_actions/session.rs	stdin (cat 'PATH' | amplihack copilot, #2660)
A	journal / distillation / merge-judge	file <key>_path — in-repo recipes (#2692/#2700)
B	overseer task_description, self-improve proposal, engineer-loop objective	bounded-inline sanitize(…,8000) (E2BIG-safe: 8000 chars ≪ ARG_MAX). File channel blocked on external recipe _path support — smart-orchestrator.yaml / simard-self-improve-cycle.yaml / simard-engineer-loop.yaml are amplihack-bundle assets (not this repo) that read {{key}} inline; filing a large value would leave the agent with an empty var. recipe_context is ready for them once those assets gain a _path read.
C	short IDs / paths / slugs (target_repo, pr_number, state_root, …)	unchanged (safe)

Tests (hermetic, > 256 KiB payloads)

tests/spawn_payload_facade.rs, e2big_argv_guard.rs, and per-path transport tests for meeting, signal, engineer (+ a no-PTY stdin round-trip), overseer-launch, self-improve. The engineer permissions test now verifies the prompt lands on stdin (not argv) while #1717 flags are preserved.

Local verification: full suite green (cargo test --tests), cargo fmt --check, cargo clippy --all-targets --all-features --locked -D warnings, mkdocs build --strict.

Docs

• Concept: docs/concepts/e2big-elimination.md
• Reference: docs/reference/large-payload-spawn-api.md
• How-to: docs/howto/add-a-safe-agent-spawn-site.md

Builds on #2660 and #2700 without regressing them.

Closes #2640

Co-authored-by: Copilot 223556219+Copilot@users.noreply.github.com

[github-actions]

📊 Coverage Summary

Generated by cargo llvm-cov --workspace --summary-only (nightly, excluding test files)

Module	Lines	Covered	Coverage
Total	146341	122010	83.4%

_{Coverage data from CI run. Test files matching tests?/ are excluded from line counts.}