Fix multiple PE/COFF parsing bugs (#1502)

[Weitao-Sun]

Fixes three PE/COFF parsing issues reported in Vector35/binaryninja#1502 (referenced there as Ghidra #9169, #9170, #9171):

• COFF symbol table OOM via unbounded symbol count and string names
• PointerToRawData not rounded to 0x200 per Windows loader behavior
• SizeOfRawData not rounded up to FileAlignment, hiding code caves

This is the api-repo counterpart of Vector35/binaryninja#1625, which will be updated to just bump the api submodule pointer once this merges.

Review feedback addressed

An earlier version of this change went through a round of review. Summary of what changed as a result:

1. virtualSize is grown to max(virtualSize, sizeOfRawData) so segments, sections, RVA characteristics, and symbol placement all agree with the corrected raw-data extent.
2. The symbol-name budget is checked before each string-table read instead of after, names are cached by string-table offset so repeated offsets are read once, and the running total accounts for the extra copies a resolved name ends up retained in.
3. The aligned raw-data size is clamped to UINT32_MAX before narrowing to uint32_t.
4. coffSymbolCount * sizeofCOFFSymbol arithmetic is widened to 64-bit in both files, and the COFF section-name-resolution path uses the runtime symbol-record size instead of a hardcoded 18.
5. Section-name and symbol-name string table offsets are validated against the declared table length before being read.
6. All six loader settings (PE + COFF, count / name-length / total-budget) now treat 0 as uncapped.
7. Magic numbers moved into named constants in peview.h, shared by both files.
8. header.coffSymbolCount is no longer mutated. Every raw symbol-table slot still gets a data variable and marker symbol, which relocation resolution looks up by address; the configurable count limit only gates the more expensive per-symbol work (name resolution, typing, aux record definitions).

Follow-up round

A second look raised four more points:

• Sector rounding was unconditional and didn't account for low-alignment PE images. Fixed in the latest commit: when SectionAlignment < page size and SectionAlignment == FileAlignment (the PE spec's low-alignment case), raw offsets map directly to RVAs without sector rounding. Checked against a synthetic low-alignment image (SectionAlignment = FileAlignment = 0x20, PointerToRawData = 0x220) — the section's file offset and content now match the declared layout instead of getting rounded down to 0x200.
• Symbol-count limits discarding valid relocations: with maxCoffSymbolCount = 1, 2 symbols, and a relocation referencing symbol index 1, the marker for index 1 is still created and the relocation resolves — this is what point 8 above covers.
• Rounded raw bytes ending up outside the mapped section: covered by the virtualSize = max(virtualSize, sizeOfRawData) change above.
• Raw-size rounding wrapping past 4 GiB to zero: the clamp to UINT32_MAX happens before narrowing (point 3 above), so this doesn't wrap.

If any of the last three still show up against this branch's current tip, a repro would help track it down further.

Known follow-up (not blocking)

The per-slot marker-symbol loop in the COFF symbol table walk (point 8) still iterates the full declared coffSymbolCount even when the count limit caps the heavier per-symbol work, so a very large declared count still costs a linear pass just to place markers. A bound derived from the file's actual size (a legitimate count can't exceed what fits in the remaining file) would close this without reintroducing the relocation-lookup issue point 8 fixed. Can follow up separately if useful.

Testing

This changes PE section-layout logic, so part of the local analysis-output test suite differs from its stored oracles — expected, since section boundaries legitimately shift for several PE test corpus binaries. The full suite gives 81 differences against this branch; running the same suite against plain dev (no commits from this branch) narrows it down:

• 23 already show up on plain dev — unrelated to this change (test_Architecture, test_memory_leaks, test_struct_return_win32/win64, PDB tests, plus a handful of non-layout tests on acpipagr.sys/acpitime.sys).
• 58 are specific to this branch, on the PE/COFF corpus binaries the change touches (acpipagr.sys, acpitime.sys, jumptable_multiple_indirect, partial_register_dataflow(_rebasing), pe_thumb, quick3dcoreplugin.dll, win32_x86.dll, win32_x86_warp.dll). Same set with or without the latest (low-alignment) commit. These are the layout-shift kind — uniform address/length changes and/or new, well-formed content in the newly-exposed byte range.

Will regenerate oracles for the 58 once the code here settles, scoped to just those binaries.

🤖 Generated with Claude Code

[Weitao-Sun]

Fixed all four in bc9cad8: name-budget now charges every retained reference including cache hits; relocations to undefined-external symbols beyond the annotation limit are resolved on demand under the same budget; RVAToFileOffset/GetRVACharacteristics now keep the last matching section on overlap; string-table reads are bounds-checked and capped to the table's remaining length in all three lookup sites in these files (also found and fixed the same missing bounds check in the section-name lookup, which wasn't flagged but has the same issue).