Gate transfer decode to exact Balances calls

[LandynDev]

decode_transfer accepted any call with transfer in its name — no pallet check. Both decode paths are now gated to exact Balances.{transfer_allow_death, transfer_keep_alive}; adds decode_stake_transfer for SubtensorModule.transfer_stake (needed by the upcoming chain watcher).

Severity: hardening, not a live hole — transfer_all/transfer_stake decoded with amount=0, and both call sites require amount >= expected, so no payment could ever have been falsely approved. The real prior defect was false negatives (e.g. a genuine transfer_all payment could never verify) plus a lookalike-call surface the exact-name gate now closes. The raw path's 7: transfer_all entry never matched anything — transfer_all is call index 4 on Subtensor — so it is dropped rather than fixed, and the rejection test pins index 4.

Before → after (structured path, same fixtures as the new tests):

transfer_all    ('0xabc', '5Miner…', 0, '5User…')    → None        # previously amount 0: a real payment that could never verify
transfer_stake  ('0xdef', '', 0, '5User…')           → None        # not a TAO transfer at all
keep_alive      ('0x123', '5Miner…', 5000, '5User…') → unchanged

Independent fix: the decode is only called from fetch_matching_tx / find_recent_outgoing in the same file, and call_module is present in the installed scalecodec serialization. Full suite: 814 passed.

[anderdc]

tests/test_decode_transfer.py:124 — transfer_all is call index 4 on Subtensor, not 7; index 7 doesn't exist in the Balances pallet (verified against live finney metadata: 0/2/3/4/5/6/8/9/10). test_raw_path_rejects_transfer_all_index passes for any unmapped index, so it pins nothing — use call_idx=4. allways/chain_providers/subtensor.py:26-28 — the raw path never decoded transfer_all either, since 4 was never in _TRANSFER_CALLS; drop that claim from the comment and the PR body so nobody later "restores" 7. The structured-path gate itself is correct and verified against real mainnet extrinsics on the pinned async-substrate-interface 2.2.1. Worth stating the true severity too: transfer_all/transfer_stake decoded with amount 0 and both call sites require amount >= expected, so no payment could ever have been falsely approved — this is hardening, not a live hole.

[LandynDev]

All three points addressed in the latest push: rejection test now pins the real index 4 (with a comment naming it so the pin is legible), the index-7/keep_alive claim is dropped from the module comment, test docstring, and PR body, and the body now states the honest severity — hardening against false negatives + lookalike calls, not a falsely-approving hole. Thanks for checking against live finney metadata.