v6.5.0

• release(6.5.0): promote v381 — #1101 comma-free fix + digit-ownership no-fragment shard (11f4bc1)
• fix(#1101): punct-drop CLOSES the US comma-free rooftop loss (v380 probe verified) + 8k ship config (295b8db)
• fix(#1101): enable augment_punct_drop_prob to close comma-free rooftop loss + track the untracked US twin (aaaa023)
• analysis(scratchpad): digit-ownership arc CONCLUSION — clean ceiling ~0.77, bar-clearing needs a trade (fb57c81)
• feat(track-b): B4b tweak 8k config (v3.7.3) — escalate the plateau-breaking all-length boost (8bf15de)
• analysis(scratchpad): B4b knob 3 tweak (v372) — the plateau BROKE; best result of the arc, clean (265e018)
• feat(track-b): B4b knob 3 tweak (v3.7.2) — maxed long-number boost (all lengths, 6x) (031e377)
• analysis(scratchpad): B4b knob 3 8k (v371) — clean additive net-positive, but target PLATEAUS below bar (ca0dc45)
• feat(track-b): B4b knob 3 8k config (v3.7.1) — escalate the additive long-number lever (36a0fce)
• analysis(scratchpad): B4b knob 3 — the first PURELY ADDITIVE lever; missed strict 2k bar, warrants 8k (e591061)
• feat(track-b): B4b knob 3 — long-number oversample (the last corpus probe) (d82bb70)
• analysis(scratchpad): B4b knob 1 (bare-street 0.30->0.70) FALSIFIED — wrong direction, target is dual-mode (54b9ff9)
• feat(track-b): B4b no-fragment knob 1 — bare-street-prob 0.30->0.70 (2c47c08)
• analysis(scratchpad): B4c 3-digit 8k (v354) — DON'T PROMOTE, the cost moved not vanished (3ff42f6)
• analysis(scratchpad): B4c 10-999 (v353) — 2-digit doesn't recover date-name; 3-digit (v352) is the cut (b2c5a0a)
• analysis(scratchpad): B4c 3-digit refinement (v352) — guard HOLDS, house_number win without postcode cost (c575902)
• analysis(scratchpad): B4c numsplice verdict — a real trade, house_number up / postcode down (22aa6bc)
• feat(train): v3.5.1-numsplice-full — B4c 8k, the numsplice probe PASSED (9034f89)
• feat(train): v3.5.0-numsplice-probe — B4c number-piece vocab splice (2k probe) (cb1671d)
• analysis(scratchpad): the Norway retrain FALSIFIED the coverage hypothesis (06f66be)
• feat(train): v3.4.1-norway-full — the 8k escalation (what v310 would've been with #1145) (0885a65)
• feat(train): v3.4.0-norway-coverage-probe — does un-dropping Norway fix the Norwegian rows? (a8b1409)
• analysis(scratchpad): the deepparse house_number rematch, re-read through Track B (655d877)
• ci(publish): fail-fast preflight when a version isn't staged on Hugging Face (f31fbe2)