[michalharakal]

Closes #1500. Part of #932.

SKEEP-003 was marked Implemented, but an audit of develop at 0.64.0 found the old storage layer still in place beside the new one. This PR changes records only; no behaviour changes.

• SKEEP-003: Status: back to Accepted, and a new Implementation status section with one table per acceptance criterion (two partly met, one not met, one without a baseline, one without a test) and one for the decisions that are open, each linked to its issue (SKEEP-003 acceptance guards: downcast guard, flat direct-memory test, benchmark baseline #1501, SKEEP-003 removals: LogicalDType, BufferHandle.Aliased, the old TensorView marker and SlicedTensorView, unused storage types #1502, SKEEP-003 decisions 3, 8 and 10: packed get() warning, axis labels on Shape, eager activation ids #1503, SKEEP-003: dispatch on Format instead of TensorData downcasts, loaders return views, TensorData as true façades #1507). Decisions 4, 5 and 7 are recorded as outside the SKEEP (scaledDotProductAttention and the backend attention kernels accept the windowed KV pair #1504, StableHLO export inserts the requant adapter a block encoding asks for #1505, Optional remote-IO module: HTTP range-request sources for SuspendingRandomAccessSource #1506); the graph.Layout deprecation is left to SKEEP-004, which already schedules it.
• SKEEP-003a: .vmfb parity acceptance run (eager vs compiled) #1148 was listed as open; it was closed as superseded, and this repository has no eager-versus-compiled parity test. Its index status Resolved is not a defined value and becomes Implemented.
• SKEEP-002: the status note said packed weights were blocked on Packed-quant byte-order (block layout) is an unwritten, contradictory contract across the engine and downstream converters #973 and that no device numbers exist. feat(#1189): packed-tensor mapped staging — Q4_K/Q6_K served straight from the mmap on Android #1190 serves Q4_K and Q6_K from the mapping, and M2-A5: a Q4_K_M model loading under a real heap cap on an Android device #1130 measured it on a Pixel 8a. The note now says that, and names what is still open (the other packed formats, the Q8_0 heap figure, windowed mapping above 2 GB). The status stays Draft.
• memory-model.adoc: the limit "packed weights still reach the managed heap" is narrowed to the formats it still applies to, and "TensorData implementations are façades over views" is worded as the code is.
• TensorStorage KDoc: no longer tells new loaders, planners and backends to target it.

🤖 Generated with Claude Code

[github-actions]

📖 Documentation Preview

The documentation has been built successfully for this PR.

Generated Files:

• Operator documentation: docs/modules/operators/_generated_/
• JSON schema output: operators.json

Artifacts:

• Download the documentation-preview-1508 artifact to view the complete documentation locally.

This comment will be updated automatically when the PR is updated.