Sub-agents cannot reliably execute: wall-time deaths lose uncommitted work, provider-route failures block dispatch, shell tooling needs workarounds

## Summary

Sub-agents (workers/builders) cannot reliably run shell commands in this environment, which makes delegated execution work — the core Fleet value proposition — unusable. Three failure modes observed today alone:

1. **Wall-time budget deaths mid-task.** Two worker sub-agents (117 and 83 steps in) hit the 1800s wall-time limit and were killed with all uncommitted work. One had a complete, correct fix uncommitted (the sub-agent transcript scroll fix); the other produced nothing. The runtime's own lessons file (GROK-UX-LESSONS-20260820.md section 0) documents the same class of failure killing research sub-agents: a scout died at 60/60 steps with no report. Budget death must produce a final report from the checkpoint and force-commit uncommitted work, not silently discard it.
2. **Provider-route brittleness blocks dispatch entirely.** Fleet profiles pinned to xAI (`builder`, `reviewer` on grok-4.6) fail at start with "client could not be built" even after `codewhale auth external-consent --provider xai --mode read-only` succeeds and saves consent — fresh processes still fail, so the consent path in the shipped 0.9.10 binary does not resolve credentials the runtime itself just persisted. There is no fallback: the task never starts.
3. **Shell execution from sub-agents is unreliable/hostile to long commands.** Background jobs get parked to /jobs with completion arriving as runtime events that may or may not surface; heredocs with certain payloads fail to spawn at all ("nul byte found in provided data"); interleaved foreground/background waits make multi-step verification (edit → revert → test → restore) awkward enough that the parent ends up doing the work itself — defeating the point of delegation.

## Expected

- A sub-agent should be a first-class executor: reliable shell foreground execution with captured exit codes, straightforward file read/write (including binary-safe edits), and predictable background-job semantics.
- Budget exhaustion = graceful landing (final report + commit of finished pieces), never silent loss.
- Route failures (missing provider creds) should fall back or queue with a clear operator-facing signal, not fail the dispatch.
- The parent's own shell tooling should not need workarounds (python-script files instead of heredocs, split greps instead of single pipelines) for routine multi-step verification.

## Evidence

- Session of 2026-08-20: two `subagent_failed` events (`wall_time_budget`, 1800s, 83 and 117 steps) — one left salvageable uncommitted WIP in `/Volumes/VIXinSSD/CW/worktrees/cw-0911-subagent-scroll-20260820`, the other nothing.
- `codewhale auth external-consent --provider xai --mode read-only --path ~/.grok/auth.json --yes` → "saved read-only external credential consent" (exit 0), yet `builder`-profile dispatch still fails with "xAI OAuth credentials not found"; `codewhale --provider xai --model grok-4.6 exec ...` in a fresh process fails the same way.
- GROK-UX-LESSONS-20260820.md documents the step-budget death pattern and recommends soft-landing + report-on-exhaustion.

## Impact

Delegation is currently a trap: dispatch succeeds, work proceeds, and then everything is lost at a hard limit or a route error. Until this is fixed, heavy lanes must stay with the parent session, which concentrates risk and burns the primary budget.

Start by reproducing the three reported failures: wall-time termination, xAI provider dispatch after external-consent, and long or background shell commands. Read GROK-UX-LESSONS-20260820.md section 0 and inspect the cited worktree for salvageable changes. Done means completed work is preserved with a final report, provider failures produce a clear fallback or queue signal, and routine shell execution is predictable.

Written by the indexing model from the issue text.