ROAD-V1-C3's named tail records DSpark as unspiked three days after it landed, and files TLI under the wrong lane

[localai-bot]

ROAD-V1-C3's recorded tail is "DSpark (SPEC-DSPARK) + heterogeneous-vocabulary
TLI (SPEC-TLI) unspiked — overlaps ROAD-V1-D3"
(.agents/specs/roadmap-v1-completion.md:40, punch-list item 17 at :216).
Every clause of that sentence is wrong, and it is wrong in two different
directions.

DSpark is not unspiked — it is implemented, gated and measured cross-engine

• Spike spec committed 2026-08-09 (2b342620e), now 1595 lines with
  W1–W8 and §§6a–6aa of measurement: .agents/specs/dspark-spec-decode.md.
• Engine matrix SPEC-DSPARK has been ACTIVE, owner CLAIM-SPEC-DSPARK,
  since that spike.
• Shipped code: qwen3_dspark.{h,cpp} + _weights.cpp,
  v1/worker/gpu/spec_decode/dspark/speculator.{h,cpp}, five test files, and
  the Speculators-format config translation.
• W7 (SPEC-DSPARK W7: the sequential Markov sample loop is HOST-side — 28% of the draft step, 14.9 MB downloaded per step #436) moved the sequential Markov sample on device; W8 (SPEC-DSPARK W8: capture the T=1+k speculative VERIFY shape (the measured 0.870x-0.981x gap) #442) captured
  the T=1+k verify by mirroring vLLM's uniform_decode_query_len = 1 + num_speculative_tokens.
• Measured against the pinned graphed oracle under pinned clocks: 35B-A3B MoE
  0.975x code cell / 1.012x prose cell, residual attributed to a 12.9%
  effective-DRAM-bandwidth gap inside marlin_moe_wna16::Marlin on
  byte-equivalent machine code. docs/BENCHMARKS.md already carries that row.

TLI is untouched, and it is not a DSpark tail at all

Upstream TLI is use_heterogeneous_vocab (config/speculative.py:150) plus
VocabMapping (v1/spec_decode/vocab_mapping.py:68). Both are consumed only
by v1/spec_decode/llm_base_proposer.py (SpecDecodeBaseProposer) and
v1/spec_decode/draft_model.py:19. The V2-runner speculators our DFlash/DSpark
port mirrors (v1/worker/gpu/spec_decode/{dflash,dspark}/) have no
heterogeneous-vocab path at all. So TLI's host row is SPEC-DRAFT-MODEL, whose
local state is a CPU propose brick with no runner construction — TLI is
prerequisite-blocked, not merely unspiked. DSpark's d2t does not cover it:
d2t is an offset table inside one tokenizer's vocab, TLI builds a string-level
intersection across tokenizer families (BPE Ġ vs SentencePiece ▁).

D3 covers none of it

.agents/specs/spec-decode-breadth-d3.md:22 puts DSpark and TLI explicitly
out of ROAD-V1-D3's scope. D3 (DONE 2026-07-27) landed ngram +
EAGLE3-blocked and covers no part of C3's tail.

Surfaces that disagree with the tree

Surface	Says
.agents/specs/roadmap-v1-completion.md:40,216	DSpark + TLI "unspiked", overlaps D3
.agents/roadmap_v1.md:196,263	"DFlash Part C + DSpark/TLI remain"; "their dedicated spikes are not written" — SPEC-DFLASH is DONE
.agents/roadmap_v1.md:397	"SPEC-DSPARK (engine matrix, INVENTORIED)"
.agents/specs/spec-decode-inventory.md:45,72	SPEC-DSPARK INVENTORIED; "no speculators-format adapter"
docs/STATUS.md:167	dspark listed among the INVENTORIED methods, contradicting the same file's own DSpark paragraph
.agents/specs/dspark-spec-decode.md Status field	"W1-W5 LANDED … W6 PARTIAL … ~2% BEHIND spec-off" (2026-08-10)
.agents/roadmap_v1.md Open issues table	#436, #442, #513 absent, though AGENTS.md requires the intake table to agree with the spec and the PR

This issue tracks reconciling those surfaces. It changes no src/, include/
or tests/ file and asserts no new measurement.