Implement member-wise reading of collections (e.g. uproot-issue-268.root) #70

🤖 AI generated content

Line item of #9: Memberwise reading of StdVector not implemented (uproot-issue-268.root). A genuine reader gap, and a prerequisite for most real TTree payloads. The failing member is a std::vector<AnalysisTree::BranchConfig> with fByteCount 1073 and version word 0x4009 (member-wise flag 0x4000 set).

Spec: Collections §4 "Member-wise". Layout:

byteCount  version|0x4000  valueVersion  count:i32  <member 0 × count> <member 1 × count> …

§4.1 gives the per-column framing table (fundamentals bare, a base class bare, TString bare, an object member framed per element, a collection member with one frame for the whole column); §4.2 covers base classes (TStreamerBase::fBaseVersion is needed here because base version words are not on disk); §4.3 the empty case (no columns at all); §5 which containers use which mode.

Current code (from the bootstrap review, #66):

• No bare-version-word path. src/rootfilespec/container.py:218-233 (StdVector.read_as; only the pair case is handled, everything else raises at :234-235) and :327-358 (the StdMap member-wise branch) both unpack(">hIi") unconditionally and assert the value-version is 0; the vector path also asserts the checksum equals the hardcoded 0xD7BED2 (the pair<double,double> of TEfficiency::fBeta_bin_params in uproot-issue38c.root). Per §4 the value-version word is read with no byte count in front, and a checksum follows only when it is ≤ 0. Evidence, with the version-9 check relaxed: collection-forms.root desynchronises (Expected position 145 but got 131, CHit.fId == 131072) because fHits is 40 0a | 00 02 | 00000002; pairs.root fails with a bare AssertionError from the same asserts.
• The checksum does not identify a pair (§8.2): resolve the value class from the member's declared type name and use the checksum only to pick a version within it.
• Column framing is uniform, the spec's is not. container.py:338-357 reads one StreamHeader per column whenever the reader is not a _FmtReader; §4.1 has six different rules (no frame for a base class or a TString, one frame per element for object members 61/62, one object slot per element for pointers 64/69, …). pairs.root exercises all six shapes.
• TStreamerBase::fBaseVersion is parsed (src/rootfilespec/bootstrap/TStreamerInfo.py:339) and never used; §4.2 says it is the only source of a base class's version inside a member-wise column.
• A recorded pair<K,V> info is ignored: TStreamerInfo.py:45-46 short-circuits pair<…> infos and src/rootfilespec/cpptype.py:52-53 always synthesises the pair. §8 says a reader must look for a recorded info first (its element classes decide the §4.1 framing) and synthesise from the template arguments only when absent. pairs.root records pair<int,string>, pair<int,vector<short>>, pair<TString,PHit*> and pair<string,int>; collections.root records none.
• The collection version word is compared to 9 — tracked separately in Collection frame version word is hardcoded to 9; files from ROOT >= 6.36 (and < 5.26) fail #99.

Test vectors: reference/root-io-spec/data/serialization/{collections,collection-forms,pairs}.root with byte-level case.toml assertions.

Severity / size: high value, L. Files: container.py, bootstrap/streamedobject.py. Strongly coupled to #67: member-wise reading is a column-wise traversal of the same element list, which the annotation-driven update_members design cannot express.

Related: #9, #67, #99, #90 (fixed arrays of collections).

Assisted-by: claude-code:claude-fable-5-1