Reorder PromptRenderer so the highest-value context lands nearest the decode boundary

[kiki830621]

Type

refactor

Problem

PromptRenderer.render emits classes in the order names → terms → phrases, so
the highest-value content sits at the front of every rendered prompt.

Whisper conditions on the prompt through the <|startofprev|> channel
immediately preceding <|startoftranscript|>, and content nearest that boundary
is widely reported to carry outsized influence. OpenAI's reference decoder makes
the same judgement structurally — when an over-long initial_prompt must be
truncated it keeps the tail:

# whisper/decoding.py, DecodingTask._get_initial_tokens
tokens = [self.tokenizer.sot_prev] + prompt_tokens[-(self.n_ctx // 2 - 1):] + tokens

So the renderer currently puts the names — the exact content a context directory
exists to inject — in the position with the least conditioning weight, in every
prompt, not only when truncation happens.

Background

#164 reversed the engine-side clamp (design D4) to keep the prefix, because the
old suffix-keeping clamp dropped names outright on overflow. That fixed the
outright-drop, but it fixed it on the content-priority axis while the recency
argument lives on the position axis. Four independent review lenses judged D4 in
#164's verify round 1; none argued for reverting it, and three independently
derived the same durable remedy, which is also what PipelineWiringTests.swift
already says in a comment but does not implement:

reorder the renderer so the highest-value items land at the tail, not re-flip
the clamp against its own priority ordering

Expected

Highest-value items land nearest the decode boundary and survive truncation.
Concretely: renderer emits phrases → terms → names (or otherwise places names
last), while the clamp keeps whichever end now holds the names.

Both mechanisms then agree, and the context-calibration spec sentence "the
clamp SHALL preserve the highest-priority items" stays satisfied.

Notes

• This changes the prompt sent for every existing Whisper transcription, so
  it deserves a measurement rather than another reasoned flip: WER/CER on
  proper-noun-bearing audio, current order vs. reordered.
• The context-calibration delta spec's worked example ("鄭澈, Che, benchmark-driven, CoreML") pins the current order and would need updating.
• Split out of context 的 token 預算應是 per-model 參數而非全域寫死 200：7 個 engine 只有 2 個吃 prompt，其餘照樣截斷且照印 injected(N) #164 deliberately: context 的 token 預算應是 per-model 參數而非全域寫死 200：7 個 engine 只有 2 個吃 prompt，其餘照樣截斷且照印 injected(N) #164 shipped the honest documentation of the
  trade-off; this is the resolution.