feat(prompts): add retry/loop-breaking rules to bu-flash system prompt

[caffeinum]

Summary

Adds an 8-line <retry_strategy> block to system_prompt_browser_use_flash.md. The rules are a condensation of guidance that already exists in the main system_prompt.md but was stripped from the slim flash variants on the assumption that the bu-2-0 fine-tune would internalize it. In production we've observed bu-2-0 looping — re-emitting the same failing action 3+ times without changing approach — particularly on auth/MFA flows.

Added rules (paraphrased)

1. Don't repeat the same action 3+ times without visible progress.
2. If the URL hasn't changed for 3+ steps and the page looks identical, you're stuck — change strategy.
3. If a click/input did nothing, don't click the same index again — pick a different element or scroll.
4. If a field's actual value differs from what you typed (page reformatted/autocompleted), wait one step before retrying.
5. If credentials are needed and not provided, stop and report — don't invent emails/passwords/codes.

Cross-reference with upstream main prompt

Equivalent guidance already exists in upstream system_prompt.md:

• L98: "if you are on the same URL for 3+ steps without meaningful progress, or the same action fails 2-3 times, try a different approach" (rules 1 & 2)
• L88: autocomplete/combobox handling (rule 4)
• L89 / L149: don't login without credentials, blocking-error check (rule 5)
• L97, L249, L264, L267: don't repeat failing actions, break loops (rules 1 & 3)

So this isn't fork-specific tuning — it's restoring upstream's own guidance into the slim flash variant where it was dropped.

Test plan

• Smoke-test bu-2-0 on a known loopy scenario (e.g. auth0 login w/ TOTP) and check that retry counts drop
• No regressions on the existing prompt test suite (pnpm test:unit)

Note

system_prompt_flash.md (the non-browser-use flash prompt) has the same gap. Out of scope for this PR; happy to follow up if maintainers want.

🤖 Generated with Claude Code