ENG-838: cap RESILIENCE_NUDGE with an honesty floor

[alecantu7]

What

Rewrites the transient RESILIENCE_NUDGE (anton/core/llm/prompts.py, injected after 2 same-tool failures) to keep its "try a workaround first" value but add an explicit honesty floor, and drops the risky closing line "Only involve the user if the problem truly requires something only they can provide."

Why

• Redundant: the always-on PROBLEM-SOLVING RESILIENCE block in CHAT_SYSTEM_PROMPT already covers the same workaround guidance with safer user-involvement wording.
• Riskiest wording in the fabrication path: that closing line is the ENG-296 "smoking gun" — it biases away from involving the user at the moment a tool fails. No observed harm today (the STUCK verdict + act_first honesty rules win — see the ENG-296 close-out), so this is defense-in-depth to lower the probabilistic tail risk.

Kept (not deleted) for its just-in-time salience.

Diff

Before: "…Before retrying the same approach or asking the user for help, try a creative workaround… Only involve the user if the problem truly requires something only they can provide."
After: "…Before retrying the same approach, try one or two genuine workarounds… If those also fail, STOP and tell the user exactly what failed and what you need — do NOT fabricate a result, claim success, or silently give up. Once real options are exhausted, asking the user is the correct move."

Safety / scope (verified)

• Opening marker "This tool has failed twice in a row." retained — two e2e tests (test_loop_safety, test_circuit_breaker_evasion) detect the nudge by it.
• test_resilience_nudge.py checks identity against the constant (rewrite-safe).
• No change to firing logic (resilience_nudge_at), the circuit-breaker escalation, or the scratchpad SIZE/TIMEOUT nudges.
• Full suite: 1204 passed (only pre-existing scratchpad-exec env failures).

Remaining (Done-when)

• Eval: add a blocked-task case to the ENG-381 harness to measure fabrication rate before/after (prove resilience isn't weakened). Not in this PR — prompt change is text-only and test-covered; flagging so it isn't lost.
• Delivery: cowork pins anton main, so reaching the desktop path also needs the anton staging→main promotion + a cowork-server uv.lock bump.

Closes ENG-838.

[alecantu7]

Self-review

Reviewed the diff + consistency against the escalation ladder and the always-on PROBLEM-SOLVING RESILIENCE block:

• Applied: changed "try one or two genuine workarounds" → "try one or two more genuine workarounds" (43cd3f7). The nudge fires after 2 failures, so "one or two more" (3-4 total) now unambiguously aligns with the persona block's "exhaust at least 2-3 approaches" — removes a possible under-persistence read.
• Noted (no change, intended): the nudge now steers toward honest "stop and tell the user" once workarounds fail (~failure 3-4), earlier than the failure-5 circuit-breaker — that's the intended direction (honesty over grinding); the circuit-breaker stays as a backstop.

Verified marker "failed twice in a row" retained, no test pins the nudge text, firing logic + scratchpad SIZE/TIMEOUT nudges unchanged. 13 targeted tests + full suite (1204) green.