docs: retract 43 tok/s projection post-D1b failure; 32 tok/s is the current ANE decode ceiling

[john-rocky]

Summary

• PR docs: retire 56 tok/s target; ANE-native value prop (power + TTFT + 43 tok/s) #78 reframed the value prop around a triad of ~1 W power, ~1 s TTFT, and ~43 tok/s decode. The 43 tok/s number came from PR spike(D1b): compute-unit split — chunk3 on .cpuAndGPU enables overlap #77's compute-unit-split spike projection (max(c1+c2+c4_ANE, c1+c3_GPU) ≈ 23 ms/step) assuming c3 and c4 could overlap.
• PR feat(pipelining): chunk3 async on .cpuAndGPU — negative result (STOP, do not merge as default) #79 empirically invalidated that assumption: the full 2-stage pipeline regressed by 24 % on all 4 prompt categories (baseline 32.8–33.2 → pipelined 24.9–25.5 tok/s on Mac Studio), plus a bit-exact failure on summary at token 50 (fp16 rounding between ANE/GPU backends of c3). Root cause is a strict c3 → c4 data dependency that leaves only a ~1 µs Swift dict-build as the within-step overlap window against ~16 ms GPU c3. Cross-step pipelining is blocked by the symmetric token-feedback edge (c3@N+1 needs c4@N).
• This PR retracts the 43 tok/s projection on main and propagates the consequence: 32 tok/s is the measured ANE decode ceiling, the decode gap vs LiteRT-LM widens from ~20 % to ~42 %, and item 27 (GPU prefill / TTFT axis) is now the single critical-path decode-adjacent lever. The UX argument (~1 W, ~1 s TTFT, GPU-free host envelope; 32 tok/s is still ~6× human read speed) carries the pitch, not decode parity.
• Retract, don't delete. All callouts cite PR feat(pipelining): chunk3 async on .cpuAndGPU — negative result (STOP, do not merge as default) #79 / commit 7c21c7b rather than rewriting the prior reasoning chain.

Files touched (one line each)

• docs/MOBILE_2K_COMPETITIVE_PLAN.md — retraction callout at top; value prop one-liner swapped to ~1 W + ~1 s TTFT (projected, item 27) + 32 tok/s (measured, current); competitive table honest about the 42 % decode gap; §"Projection basis" 43 tok/s subsection rewritten as 32 tok/s ceiling (measured) with root-cause analysis; execution table collapses from A+B to B (item 27) only.
• docs/PHASE_B_DECISION.md §"What this means for the go-forward target" — D1b item flipped to REGRESSED with structural cause (c3→c4 data dep); item 27 called out as sole tractable decode-adjacent lever; the 43 tok/s claim explicitly retracted inline.
• docs/PRIORITY_ROADMAP.md — item 27 footnote added marking it as the single critical-path decode-adjacent item on the roadmap after D1b invalidation.
• docs/HANDOFF.md — read-order now includes the D1b failure evidence (PR feat(pipelining): chunk3 async on .cpuAndGPU — negative result (STOP, do not merge as default) #79 + PHASE_D_PIPELINING_IMPL.md on branch); opening prompt retracts 43 tok/s; next-session start options are item 27 OR one of PR feat(pipelining): chunk3 async on .cpuAndGPU — negative result (STOP, do not merge as default) #79's three conversion/-side options (decoupled c4 / speculative h3 / model re-chunking).

Net: +182 / −83 across 4 docs (~99 net lines, under the 150 cap).

Scope

• Docs only. Zero code.
• Does not touch finding docs (PHASE_B_V3/V4, PHASE_C_, BASELINE_) — historical reasoning preserved.
• Note: task item 3 asked for a retraction header on docs/PHASE_D_COMPUTE_UNIT_SPLIT_SPIKE.md (PR spike(D1b): compute-unit split — chunk3 on .cpuAndGPU enables overlap #77's spike doc). That file is not on main — PR spike(D1b): compute-unit split — chunk3 on .cpuAndGPU enables overlap #77 is still OPEN and has not merged — so there is nothing to edit on main. If/when spike(D1b): compute-unit split — chunk3 on .cpuAndGPU enables overlap #77 lands, a follow-up one-liner can add the header there.

Test plan

• Diff reads honestly: decode gap is stated as 42 %, not massaged.
• No strategic decisions reopened — only the numerical claim changes.
• Every retracted claim cites PR feat(pipelining): chunk3 async on .cpuAndGPU — negative result (STOP, do not merge as default) #79 / commit 7c21c7b.
• Commit author is John Rocky; no "claude" in message or committer.