docs(bench): speculative_bench module docs contradict its own code in three places

Start in src/bin/speculative_bench.rs by comparing the module docs and long_about with REACHABLE_PAIRINGS and the MTP, DFlash, and batch-sweep descriptions. Update the three contradictory passages so they match the code and pairing table; no run is required.

Summary

Correct three passages in speculative_bench's module docs and long_about that contradict the code in the same file: a stale target count, a batch-sweep claim the code disowns, and a "deferred" claim about a path that already records real numbers.

Background

The doc comments were written against an earlier version of the file and drifted as the pairing table and the MTP path grew. A reader trusting the header comes away with the wrong set of reachable targets, the wrong batch behavior, and the wrong idea that MTP is not yet measured.

Proposed Solution

Edit the three passages so the module docs, long_about, and the pairing table describe what the code does. No run required.

Implementation Notes

• src/bin/speculative_bench.rs:21 claims the two reachable targets models/qwen3.5-4b-4bit and models/gemma-4-31b-it-4bit; REACHABLE_PAIRINGS actually lists four targets, adding gemma-4-12b-it-4bit (:279) and qwen3.8-27b-4bit (:297).
• :61 says "Full sweep across pairings (B = 1, 2, 4)", but the sweep logs benching at B=1 and --batch's own doc notes the baseline path always runs B=1 today.
• :118-121 long_about says speculative paths are "scaffolded but deferred to follow-up", which contradicts the module body: MTP records real decode tok/s; only DFlash is deferred.
• Distinct from #1667, the --prompt panic in the same binary.

Acceptance Criteria

• Module docs, long_about, and the pairing table tell one consistent story matching the code.

Original Suggestion

Title: docs(bench): speculative_bench module docs contradict its own code in three places

speculative_bench's module docs and long_about contradict the code in the same file in three places — stale target counts, a batch-sweep claim the code disowns, and a "deferred" claim about a path that ships.

Evidence

• src/bin/speculative_bench.rs:21 — claims "the two reachable target models (models/qwen3.5-4b-4bit, models/gemma-4-31b-it-4bit)"; REACHABLE_PAIRINGS (:243-308) lists four targets, adding gemma-4-12b-it-4bit (:281) and qwen3.8-27b-4bit (:298)
• :61 — "Full sweep across pairings (B = 1, 2, 4)"; the sweep logs "benching {} pairings at B=1" (:948-950) and --batch's own doc says "The baseline path always runs B=1 today" (:139-141)
• :118-121 — long_about says "Speculative paths are scaffolded but deferred to follow-up", contradicting :26-33 and :94-99 (MTP records real decode tok/s; only DFlash at :100-103 is deferred)

Suggested fix

Update the three passages against the code in the same file. No run required.

(Distinct from #1667, the --prompt panic in this binary.)

Acceptance criteria

• Module docs, long_about, and the pairing table tell one consistent story