Reject CBOR break codes outside indefinite-length items

[gaoflow]

Fixes #305.

Problem

A break code (major type 7, subtype 31 / 0xff) is only well-formed as the terminator of an indefinite-length item. When it appears anywhere a data item is expected, the decoder leaks the internal break-marker sentinel to the caller (or accepts it as a real element) instead of rejecting the input:

>>> import cbor2
>>> cbor2.loads(bytes.fromhex('ff'))                          # naked break
<object object at 0x...>                                      # should raise
>>> cbor2.loads(bytes.fromhex('ff'), allow_indefinite=False)
<object object at 0x...>                                      # should raise
>>> cbor2.loads(bytes.fromhex('8301ff02'))                    # break inside a definite array
[1, <object object at 0x...>, 2]                              # should raise

The same leak occurs for definite-length map keys/values (a1ff01, a101ff) and for the content of a tag, shareable value or string-reference namespace (d818ff, d81cff, d90100ff, d9d9f7ff). RFC 8949 sect. 3.2.1 defines such input as not well-formed, and as noted in the issue, "under no circumstances should a naked break marker come out of the decoding process".

Cause

The break marker is produced by decode_special as an ordinary decoded value and flows through the generic frame machinery. Only the indefinite-length array/map callbacks check item.is(break_marker) to terminate; every other consumer (definite-length containers, tags, sets, shareables, string namespaces, and the top-level result) silently passes it through.

Fix

Establish the invariant that a break marker may only be consumed by an indefinite-length container. A new accepts_break flag on StackFrame is set only for indefinite-length array and map frames (indefinite-length strings consume the break internally and never produce a free value). The decode loop then rejects a break marker that reaches any other frame, and one left as the top-level result, with a CBORDecodeError.

Tests

Added test_break_outside_indefinite (12 positions: naked, before-data, definite array element/first/nested, definite map key/value, tag content, shareable, string namespace, self-describe tag, set), test_break_outside_indefinite_disabled, and test_break_terminates_indefinite (5 cases confirming real indefinite arrays/maps, nesting, and indefinite-in-definite still decode). test_string_oversized's fuzz payload embeds a break inside a definite map and now raises the break error before its truncated tail is reached, so its expectation was updated (premature-end-of-stream remains covered by other tests). Full suite passes (462).

I worked on this with AI assistance under my direction and reviewed the change myself.

[coveralls]

coverage: 94.952% (+0.06%) from 94.889% — gaoflow:fix-305-break-marker-leak into agronholm:master