put() zeroed the caller's buffer only on the success path, so a host
Subtle that rejects the encrypt left the plaintext behind in the array it
was handed. Wipe it in a finally, matching the guarantee use() already
makes for decrypted material.

The vault mints its AES-GCM master key itself and stores it beside the
material it seals, so a copy of the profile directory opens that material
with no secret. Add a `masterKey` provider to the IndexedDB driver, to
createWebKeyStore and to the WithKeyStore extension, for callers that want
the key bound to something the browser cannot produce alone — derived from
a user password, or held by another context.

The provider is resolved per sealing or opening operation rather than
captured in `ready`, so it may reject while a vault is locked and resolve
once it is open. With one set the vault mints no key of its own, `clear()`
stops preserving a record that seals nothing this driver writes, and the
driver reports nativeCryptoKey: false so that keys which would otherwise
persist as non-extractable CryptoKeys are sealed with the supplied key too.

Adopting it is a migration, not a flag, and the docs say so: material the
default vault already sealed stops opening, since `use()` asks the provider
for a key that did not seal it. Sealing standard keys also means their bytes
are decrypted into JS memory for each use. Both limits are pinned by tests.

## [2026.37.0](https://github.com/michaeltchuang/a-day-in-my-bobalife/compare/v2026.36.0...v2026.37.0) (2026-09-21)

### Updates & Maintenance

* log greeting (2026-09-15) ([5587dee](https://github.com/michaeltchuang/a-day-in-my-bobalife/commit/5587deeeebce00e038e465be6b76c91f818d551e))
* log greeting (2026-09-16) ([bec2a77](https://github.com/michaeltchuang/a-day-in-my-bobalife/commit/bec2a7756bd074e105daa2501c0bef1e277780e6))
* log greeting (2026-09-17) ([bd03815](https://github.com/michaeltchuang/a-day-in-my-bobalife/commit/bd038156bc8c7948258e1df9de00449b226cc34a))
* log greeting (2026-09-18) ([775c817](https://github.com/michaeltchuang/a-day-in-my-bobalife/commit/775c8171c55b1223e06181cea44413781f95252d))
* log greeting (2026-09-19) ([b31bd4a](https://github.com/michaeltchuang/a-day-in-my-bobalife/commit/b31bd4a6b28ed7d1aa7140ec384d10b43fd05134))
* log greeting (2026-09-20) ([eb3c3e0](https://github.com/michaeltchuang/a-day-in-my-bobalife/commit/eb3c3e04a5fa2dad1f6bd1f37087e9dc756fedb1))
* log greeting (2026-09-21) ([f1db816](https://github.com/michaeltchuang/a-day-in-my-bobalife/commit/f1db81688487e1e6e8f850f7ee2ed70e701f89f7))

Every same-origin context that builds an IndexedDB driver reaches
getMasterKey concurrently on first use. All of them missed the read, all
generated a key, the last put won, and each context kept using the key it
had generated itself, so material sealed by one could not be opened by
another — surfacing as a bare AES-GCM OperationError.

Mint the key under a Web Lock and re-read once the lock is held, so the
first caller creates and the rest reuse what it wrote.

LogicSig args are not signed so any relay can append args to a
transaction in flight. The program cannot detect it: it can prove that
at least N args exist, never that at most N do.

That is harmless while args are free. Once they are priced, and
combined with a refundable fee, padding would start consuming the
sender's extra fees.

Under the new RequireLogicSigArgAccess consensus parameter, a LogicSig may
carry no argument it did not read: none above the highest index it read,
and none below that index unless the argument is empty. Empty holes remain
legal, so goal tealsign's padding keeps working.

opArgN records reads in a bitmask on EvalContext, and logicSigVerify
checks it -- this is a rule about a transaction, not about what a program
means, so it sits beside the other transaction-level LogicSig rules.