Lane I — Moonshot bitnet.cpp bridge

Field	Value
Branch	lane/i-bitnet
Base	main @ 5910978
Status	Delivered (recipe + pin; no giant submodule)
Date	2026-08-04

Owned paths touched

• scripts/bridges/llamacpp_bitnet_recipe.py — hardened: pins, --check, --probe
• scripts/bridges/llamacpp_bitnet_pins.json — thin SHA/tag pin (preferred over submodule)
• docs/23_BITNET_CPP_BRIDGE.md — full pinned recipe + policy
• tests/test_llamacpp_bitnet_bridge.py — schema / recipe / probe smoke
• results/bridge_cpu_llamacpp_bitnet.json — regenerated from recipe
• third_party/BITNET_PIN.md — explicit non-vendor note

ROADMAP checkbox proposals (integrator only)

Do not edit twin ROADMAP from this lane. When merging, apply:

ID	Current	Propose	Evidence
W4.T06	[~] docs + bridges	[x] or keep [~] with “pinned bridge shipped” note	docs/23 + pins + tests
WC-P2	bridges first-class	still [~] until Lane E bnn bridge CLI	recipe is first-class; CLI is E
Scorecard “Bridges GPU/BitNet”	[~]	keep [~] pending CLI; handoff clarity improved	docs/23
Moonshot “bitnet.cpp submodule”	deferred	CLOSED-BY-POLICY — recipe+pin replaces submodule	pins vendor_submodule: false

Acceptance

• [x] Recipe prints pinned clone/build/quantize steps
• [x] --check validates pins without network
• [x] No microsoft/BitNet git submodule added
• [x] Docs/23 documents pins, non-goals, dual metrics
• [x] Pytest covers pins + recipe emission

Residuals

1. Lane E: first-class bnn bridge … CLI over this script (ownership: E).
2. Full upstream build not run in this lane (toolchain + multi-GB models); operators clone at pin locally.
3. Re-pin cadence: when microsoft/BitNet moves, bump llamacpp_bitnet_pins.json + docs/23 snapshot table in one PR.
4. Integrator updates docs/MOONSHOT_DEFERRALS.md row for bitnet submodule → CLOSED-BY-POLICY / recipe+pin.