The Note field doc hardcoded 1024 bytes, which went stale when v42
raised MaxAbsoluteTxnNoteBytes to 4096. Build it with Sprintf from
bounds.MaxTxnNoteBytes, the same cross-version maximum the decoder
uses, so the generated docs cannot drift again. Unlike the limits
derived from evaluator constants, this changes generated output:
1024 becomes 4096 everywhere the Note doc appears.

These files are hand-maintained, not generated, and had drifted
from the implementation:

- tutorial.md described the pre-v41 LogicSig world. It now covers
  the four signature methods including post-quantum signatures,
  the LMsig and PQsig LogicSig fields and how they bind the
  authorizer's address into the signed message, that at most one
  delegation signature may be present, that validation is against
  the authorizer (AuthAddr when rekeyed), and that legacy Msig is
  rejected from consensus v41 onward, with goal's --legacy-msig
  escape hatch. Also fixes the "LogicSic" typo.
- jsonspec.md conflated rules that json_ref enforces while parsing
  the whole text with rules enforced only on the extracted value.
  Classify each rule as whole-text or extraction. In particular,
  the old "Raw non-unicode characters not accepted" claim was
  wrong: invalid utf-8 parses, and becomes U+FFFD when a string
  value containing it is extracted. A \u escape with non-hex
  digits parses but errors on extraction (it becomes U+FFFD only
  in a key), and a null value parses but always errors on
  extraction.
- TestParseRawNonUnicodeChar never tested invalid utf-8 (its input
  was valid utf-8 and its error cases failed for other reasons).
  It now feeds a genuine invalid byte through parse and extraction
  and asserts the U+FFFD replacement.
- README.md points at the generated stack-types table and lists
  the assembler pseudo-ops (int, byte, addr, method) that appear
  in no opcode table.

Document behavior that the descriptions omitted or got wrong:

- The v13 box opcodes: shared boxAccessExtra text states the name
  length rule (1-64 bytes, MaxAppKeyLen), the availability rule,
  and the ClearState ban for every box opcode, not only those
  whose short description mentioned them. foreignBoxExtra spells
  out the app_box_* authorization matrix (AppForeignBoxReads for
  reads, AppFamilyBoxAccess plus same creator for everything
  else) and the family reentrancy rule; boxFamilyExtra applies the
  same reentrancy rule to an app's own boxes once it opts in.
  app_params_set explains what setting each flag grants.
- itxn_begin's Fee description matched neither the group-based fee
  logic nor when fees are checked. Describe the actual default:
  the amount that would make the group's fees sufficient so far,
  with the new transaction counted as one base fee, and note that
  the real check happens at itxn_submit.
- bnz documents the v13 varint offset encoding (previously it
  described only the two-byte form), including the forward/back
  reference points and that a branch to the start of its own
  instruction cannot be encoded. callsub points at bnz for the
  encoding.
- State failure conditions: shl/shr fail if B > 63, divmodw fails
  on a zero divisor, bzero fails above 4096.
- balance/min_balance params claimed an application id argument
  they do not take. voter_params_get and online_stake answer as of
  the balance round (320 rounds back), not the agreement round.
- json_ref links to the JSON spec by URL (the relative link only
  worked in the repo) and assorted typo/spacing fixes, including
  the ec_multi_scalar_mul naming note that compared the opcode to
  itself.

Deriving docVersion from the current consensus LogicSigVersion meant
new opcodes could not be documented or checked in CI until their
consensus version was released. Make it a manually bumped package
constant again, as it was before, so in-progress opcodes can be
documented during development. The LogicSigVersion field in the
langspec files stays consensus-derived, since it reports what the
chain currently accepts, which can lag docVersion.

Generated output is unchanged.