The original five small tools carried the 2026-08-16 session; the directory now also contains later campaign diagnostics. Unless a tool says otherwise, run it from the worktree root after ninja has produced build/base/** and build/split/**.

Hard-coded path warning: units_hunt_all.py and tools/c2dbg32/*.py embed the absolute path C:\halo-worktrees\claude-finish-hs-20260816\xbox\bin\vc7\CL.Exe. If you work in a different worktree, edit the CL constant in each. gate.py uses HALO_CL when set, then the current worktree's xbox/bin/vc7/CL.Exe, and retains the old absolute path only as a compatibility fallback. Everything else is path-relative.

board.py — the scoreboard

python tools/campaign/board.py                 # totals + selected small partial objects
python tools/campaign/board.py source/hs/hs    # one unit, listing its non-exact functions

Walks objdiff.json and gates the eligible _-prefixed .text function-start owners with coff_compare.section_infos_equal. Only EXACT rows contribute to its done totals. Ordinary fuzzy PARKED rows do not count; asm-implemented and vendored-assembly rows are displayed separately as ASM and do not count in this C-reconstruction board even when their bytes are equal.

Board byte sizes include target section padding. Its selected function-owner population and function-only "fully exact" objects are not the canonical meaningful-byte or whole-object completion totals. Use the semantic-adjusted production progress report for meaningful credited bytes, and require the data/BSS/linkage/extra-owner and source-authenticity admission checks before marking an object Matching.

The default per-object listing is filtered to partial objects with some exact work and at most 8,000 remaining padded bytes. It is not an exhaustive census and must not be diffed as a regression proof; use stable_verdicts.py for the complete built-owner set. To inspect objects within N functions of completing their listed function set:

python tools/campaign/board.py | grep -E 'remaining +[0-9]+ B / [123] fn'

gate.py — compile-and-gate one unit, optionally with edits

python tools/campaign/gate.py source/units/units                       # whole unit
python tools/campaign/gate.py source/units/units --fn _unit_place      # one function
python tools/campaign/gate.py source/units/units --edits e.json --fn _unit_place --disas _unit_place
python tools/campaign/gate.py source/units/units --cflag /Ob1 --all    # diagnostic flag override
python tools/campaign/gate.py source/ai/encounters --alias old_name=new_name --all
python tools/campaign/gate.py source/ai/encounters --out scratch/encounters-probe.obj
python tools/campaign/gate.py source/physics/collision_bsp --forbid-emitted-symbol _point_from_line3d --all

--edits takes a JSON list of [find, replace, tag] triples applied to the real source in memory (the file on disk is never touched), so you can sweep spellings without dirtying the tree. --disas prints a side-by-side target/ours disassembly with relocation annotations, but pairs instructions by index rather than aligning insertions/deletions. Use alndiff.py for actual instruction alignment. Spelling sweeps are not source-authenticity evidence and are not authorized by the current house rules merely because the tool supports them. Reads per-unit cflags out of build.ninja, and always prepends /nologo /c (libcmt units omit them). --cflag appends one diagnostic compiler option after those unit flags and may be repeated. It is an audit aid only: a flag improvement is not source-match evidence and must not be admitted without independent build provenance. --alias applies a whole-identifier source rename in memory for symbol-mapping probes; like --edits, it never changes the source file and is not admission evidence by itself. --out preserves the PID-isolated probe object at the specified path for independent COFF or disassembly inspection. --all prints exact functions as well as residual and unwritten functions.

--forbid-emitted-symbol NAME checks the complete candidate COFF symbol table for a code definition with that exact decorated name and exits nonzero if one is present. Repeat the option to forbid additional names. The check includes candidate-only functions absent from the target, regardless of --fn, as well as static definitions, named code subsections, and nonzero symbol offsets. Undefined references, COMMON symbols, and data definitions do not fail this code-emission check. Failures identify the object, symbol-table index, section, offset, and section size for inspection. A clean check prints an explicit pass; without this option the existing gate output and exit behavior are unchanged. Use this guard to verify the January point_from_line3d inline schedule: the ordinary target-function listing cannot detect an extra candidate-only helper.

apply_edits.py — land an edits JSON on the real file

python tools/campaign/apply_edits.py e.json source/units/units.c

Normalizes LF patterns to the file's CRLF. Write edit JSON with a script, not a bash heredoc — heredocs eat a backslash level and silently corrupt "c:\\halo\\SOURCE\\..." assert paths (cost real time this session; use chr(92) if you must build such a string in Python).

alias_scan.py — the naming-gap scanner

python tools/campaign/alias_scan.py

Finds target _code_XXXXXXXX sections whose size + normalized sha equal a section our base object already emits under a real name. Rename the config/symbols.json entry at that file_offset, run ninja (csplit regenerates the split), and the pair gates EXACT. Closed two whole objects this session. Skip candidates that match multiple our-names — ambiguous.

stable_verdicts.py — rename-stable whole-tree regression snapshots

python -m tools.campaign.stable_verdicts snapshot scratch/before.json
python -m tools.campaign.stable_verdicts snapshot scratch/after.json
python -m tools.campaign.stable_verdicts diff scratch/before.json scratch/after.json

Keys each function by unit plus January target section index, so a reviewed semantic symbol rename does not appear as one lost function and one new function. The diff exits nonzero for every exact-to-residual/unwritten loss and reports the padded bytes gained.

alndiff.py — aligned normalized disassembly differences

python tools/campaign/alndiff.py source/objects/damage _object_damage_update
python tools/campaign/alndiff.py source/ai/actors _code_00123456 --ours-function _actor_helper

Disassembles exactly one target and candidate function section with Capstone, masks relocated dwords, preserves relocation identities as annotations, and aligns the normalized instruction streams with difflib.SequenceMatcher. Unlike an index-aligned zip, an inserted block does not make the rest of the function look different. --target-object and --ours-object inspect explicit objects; relative overrides are resolved from --root.

tinfo.py — bounded read-only COFF inspection

python tools/campaign/tinfo.py source/objects/damage
python tools/campaign/tinfo.py source/objects/damage --fn _object_damage_update
python tools/campaign/tinfo.py source/objects/damage --data --preview-bytes 128
python tools/campaign/tinfo.py --object scratch/probe.obj --json

Lists function section sizes, normalized hashes, and relocation identities, or prints bounded previews of non-code sections. It opens objects read-only and never materializes uninitialized/BSS contents. The default object is the January split target; pass --base for the built candidate.

namegap.py — proposal-only symbol-name gaps

python tools/campaign/namegap.py source/objects/damage
python tools/campaign/namegap.py --json

Finds unmatched target/candidate function sections that are strictly equal under coff_compare.section_infos_equal and uniquely paired in both directions. Results are non-authoritative proposals, not original-name proof. The tool never edits source or config, rejects ambiguous pairings, and requires independent January/PDB/source provenance review before a human makes any rename.

merge_candidates.py — fail-closed candidate-copy merger

python tools/campaign/merge_candidates.py scratch/base.c scratch/merged.c scratch/a.c scratch/b.c

Merges disjoint edits from complete candidate copies made against one common base. It rejects overlapping replacements, same-position insertions, mixed or changed line endings, and output paths outside scratch/. Output is written atomically and remains a review artifact; diff and gate every affected function before applying it to source.

relocdiff.py — complete relocation-identity comparison

python tools/campaign/relocdiff.py source/hs/hs_compile _hs_parse_object_name
python tools/campaign/relocdiff.py source/hs/hs_compile _hs_parse_object_name scratch/probe.obj --count-by-target

Prints both sides of every relocation row, including address, type, target identity, unmatched tails, and optional per-target counts. The default mode fails closed unless size, relocation count, normalized bytes, and address/type layout establish a pure relocation-identity residual. Use --allow-structural only for deliberate broader inspection.

stdcall_scan.py — proposal-only calling-convention name gaps

python tools/campaign/stdcall_scan.py source/hs/hs_compile
python tools/campaign/stdcall_scan.py --all

Scans terminal x86 ret-immediate instructions and proposes missing stdcall decorations without editing config. By default it reports only _code_XXXXXXXX placeholders; --all-undecorated broadens the diagnostic. Every hit remains non-authoritative and requires source, symbol, and caller review. Missing units fail nonzero, while --keep-going permits an explicitly partial scan that still exits nonzero.

units_hunt_all.py — cross-branch splice hunter

python tools/campaign/units_hunt_all.py _unit_update _unit_died

Collects every unique version of the named functions across all ~1,100 local branches' source/units/units.c, splices each into the current file, compiles, and gates. Finds work already done in another lane. Generalize by editing the unit path constants.

branch_sweep.py — fast historical source census

python tools/campaign/branch_sweep.py source/ai/actor_type_infection
python tools/campaign/branch_sweep.py --near

Compiles each distinct reachable historical blob for the selected translation unit against the current target and flags versions that gate better than the current source. Blob discovery uses one git rev-list --objects walk per file, so runtime scales with distinct source versions rather than with the repository's large branch count. Map a promising history:<blob> result back to commits with git log --all --find-object=<blob> before source and provenance review.

tools/bijection_walk.py — register-bijection lockstep walker; the instrument that separates real structural divergence from allocation permutation. python tools/bijection_walk.py --unit source/units/units _unit_enter_seat -v
tools/coff_compare.py — the comparator; section_infos_equal is the gate.
tools/c2dbg32/ — the 32-bit C2.Dll debugger (see its own README).

README.md

tools/campaign — the working harnesses