[ashhart]

Release prep, docs only. An independent check of main's Unreleased section against the 104 PRs merged since v1.2.1, applied byte-exact:

• 12 lines gain their PR number (they named only the issue);
• 2 wrong PR numbers are fixed, and dsv41 metal: bf16 and fp32 matvecs and the phase-1 glue #1710/dsv41 metal: the registered glue kernels (norms, RoPE, FP8 and FP4 round trips, embed, join, mean) #1711 now have their line;
• 3 duplicates go: drama3 cuda: reuse a voice reference's prefill across requests #1745's two Drama lines, Lane host: opt-in FCFS chunked admission for long concurrent prompts (flashnext Metal) #1843 and the --slide refusal (both already in 1.2.1);
• qwen3_5 cuda: deferred keeps opt-in until their copy's INVALID_VALUE is fixed (#1824) #1958/perf(qwen3_5 cuda): kept DeltaNet rows replay in the next slot scan instead of a full-state keep (+1 to +8 %, same bits) #1824's contradiction becomes one opt-in line;
• 3 figures are fixed: hip: decode router over the bf16 rows the checkpoint holds #1614's BF16 router costs 2.5 MiB a layer rather than saving it; flashnext cuda: experts load past the page cache with core's direct_gather, and a load keeps its shards open (cold load 74 → 48 s) #1901's cold start is 74 to 48 s; perf(qwen3_5 cuda): kept DeltaNet rows replay in the next slot scan instead of a full-state keep (+1 to +8 %, same bits) #1824 is +1 to +8%;
• glm cuda: a prompt chunk's MTP pass runs its last row past the caches #1577's and exl3 cuda: the 27B's EXL3 speed series on the native Dense module (#260), stacked on #1465 #1467's figures now match their receipts;
• 6 missing blank lines are added.

No code changes.