[ROADMAP][EMBER-07] Seven-billion-parameter hard rung

[wordingone]

Canonical contract: https://github.com/wordingone/ember/blob/master/docs/roadmap/milestones/EMBER-07.md

This issue tracks concrete work and evidence. Its progress does not complete or redefine the version-controlled milestone contract.

No child issue may be closed without current-master evidence for its own acceptance clauses.

[wordingone]

HISTORICAL #29 SCOPE ROUTING — no 7B scale result is claimed.

The complete lossless C-SCALE transfer and W1/W2 negative-evidence boundary are recorded here:

#1119 (comment)

EMBER-07 inherits the first later-rung subset: preregister the >=7B target and predicted widening signature; compare against the frozen 3B checkpoint plus matched dense restart, predetermined schedule growth, and ungated accumulation at equal tokens/FLOPs; measure exact trained and active capacity, GPU memory, wall/energy/bytes/checkpoint/cumulative cost; require deletion-attributed learned gains; and keep active cost inside the current 24GB envelope. Historical cbase growth is not lineage credit.

Current contract: docs/roadmap/milestones/EMBER-07.md blob 5807569c02d36cfbac855327129920b526e83f8c.

Review provenance: KAI_SOLO, NOT_CROSS_REVIEWED.

[wordingone]

Intermediate-scale scope absorbed from historical issue #241

For this canonical 7B hard-rung parent, any intermediate candidate near the historical 4.24B total-parameter shape is an engineering measurement on the path to 7B, not a completed rung or promotion.

The following #241 clauses are therefore explicit here:

• before an intermediate candidate can inform the 7B launch, measure production-path peak GPU memory and publish a per-term ledger, including optimizer/update state and transient gradient storage;
• test per-layer gradient fusion and bf16 momentum as evidence-backed candidate mechanisms where the live cost model warrants them, without making either historical cbase implementation mandatory;
• require a matched loss-trajectory control before promoting the mechanism;
• enforce this milestone's existing bounds: peak device memory at most 23.5 GiB, active trainable capacity at most 4B per episode, and matched per-token cost upper bounds at most 1.25 times the frozen 3B reference;
• do not restore predetermined cbase rung scheduling. Growth remains differentiated verified experience, one candidate expert family at a time.

The historical B0 int8-momentum result remains a rejection, not a promotable 14.6 GiB path. Nothing in this transfer claims a 4.24B or 7B checkpoint, sufficient training, memory fit, or capability.

refs #241 #869 #207 #591