[BUG] Inline prompt suggestions silently suppressed while rate-limit status is allowed_warning (returns at allowed)

Preflight Checklist

• [x] I have searched existing issues and this hasn't been reported yet
• [x] This is a single bug report
• [x] I am using the latest version of Claude Code

What's Wrong?

Inline prompt suggestions (the dimmed "ghost text" next-step hint in the input
box) stop appearing once my account is near its weekly usage limit, and return
the moment usage resets — with nothing else changed.

The correlation is exact and reproducible against the public
anthropic-ratelimit-unified-status response header:

• When the header reads allowed → ghost-text suggestions appear normally.
• When the header reads allowed_warning (the near-limit state) → suggestions

are fully suppressed for the entire window, even though every normal request
still succeeds and the session is completely usable.

• On the weekly window reset, the header flips back to allowed and suggestions

return immediately on the next idle turn.

This persists even with CLAUDE_CODE_ENABLE_PROMPT_SUGGESTION=1 set, so the env
var is not a workaround for this particular case.

Why this looks like a bug, not intended throttling

The allowed_warning state is still fully usable — requests succeed — so
suppressing an ambient UI hint for the whole window (rather than only on actual
exhaustion) seems unintended. It also has no user-visible signal, so people
misattribute it to the older March regression (#30699) and chase the env-var /
budgetTokens fixes that don't apply here.

This appears to be the same strict-equality class of defect already fixed in
#57822 ("promptSuggestion silently disabled … due to strict equality mismatch
on querySource") — there, the gate failed for a valid non-default value; here it
looks like the suggestion gate treats anything other than exactly allowed as
off, so the near-limit allowed_warning state disables it.

What Should Happen?

allowed_warning is a still-usable state, so suggestions should remain available
in it (suppress only on real exhaustion), or this should be a documented, user-
controllable behavior.

Steps to Reproduce

1. Use an account until the weekly usage window crosses the warning threshold

(the anthropic-ratelimit-unified-status response header reads allowed_warning).

1. Have a normal multi-turn conversation; wait idle after a response.
2. Observe: no ghost-text suggestion appears, despite requests succeeding and

CLAUDE_CODE_ENABLE_PROMPT_SUGGESTION=1.

1. After the weekly reset (header back to allowed), repeat — suggestions return.

Environment

• Claude Code version: 2.1.195
• OS: Linux
• Related: #57822 (same defect class, fixed), #30699 (separate March regression

often conflated with this)