fix(history): judge the pushed range, not the whole tip tree - #627

How I verified this

5 commands recorded - as recorded (shortened, folded onto one line), grouped by kind. No entry asserts a pass or a fail: read the output. Not necessarily everything the session ran.

test

• git log --oneline -1 && echo "---" && uv run pytest -q "tests/test_codex_oversized_jsonl_line.py::test_boundary_sizes_all_round_trip_byte_exact[1048576]" 2>&1 | tail -60

4c2bdd7c0 [WIP-BLOCKED] The history gate reports the tree, not the range it was asked about
---
warning: `VIRTUAL_ENV=/Users/eyalgolan/git/<redacted>-public/.venv` does not match the project environment path `.venv` and will be ignored; use `--active` to target the active environment instead
.                                                                        [100%]
1 passed in 1.73s

• uv run pytest -q tests/test_history_gate_range_attribution.py tests/test_history_gate_hit_report.py 2>&1 | tail -100

warning: `VIRTUAL_ENV=/Users/eyalgolan/git/<redacted>-public/.venv` does not match the project environment path `.venv` and will be ignored; use `--active` to target the active environment instead
.................................                                        [100%]
33 passed in 18.12s

• uv run pytest -q -s tests/test_history_gate_range_attribution.py::test_pre_push_hook_path_emits_the_full_hit_list 2>&1 | tail -100

warning: `VIRTUAL_ENV=/Users/eyalgolan/git/<redacted>-public/.venv` does not match the project environment path `.venv` and will be ignored; use `--active` to target the active environment instead
.
1 passed in 1.68s

• cp scripts/history_gate_hit_report.py /tmp/current_history_gate_backup.py cp /tmp/base_history_gate.py scripts/history_gate_hit_report.py echo "--- swapped to pre-fix baseline ---" uv run pytest -q tests/t [... 235 of 578 characters omitted from the middle ...] ain scripts/history_gate_hit_report.py diff -q /tmp/current_history_gate_backup.py scripts/history_gate_hit_report.py && echo "RESTORE OK"

--- swapped to pre-fix baseline ---
warning: `VIRTUAL_ENV=/Users/eyalgolan/git/<redacted>-public/.venv` does not match the project environment path `.venv` and will be ignored; use `--active` to target the active environment instead
F                                                                        [100%]
=================================== FAILURES ===================================
________ test_range_verdict_reports_a_hit_planted_in_the_scanned_range _________

scratch_repo = {'repo': PosixPath('/private/var/folders/1r/3r0rt1jd4j1456rsg_fh4d380000gn/T/pytest-of-eyalgolan/pytest-1391/test_rang...0145f791f3a43a', 'B': 'ddec836f87831367858953142cce3fa9c740150b', 'C': 
[... 589 of 1,728 characters omitted from the middle ...]
 rc = hgr.cmd_gate(args)
             ^^^^^^^^^^^^
E       AttributeError: module '_nh_history_gate_range_attribution' has no attribute 'cmd_gate'

tests/test_history_gate_range_attribution.py:241: AttributeError
=========================== short test summary info ============================
FAILED tests/test_history_gate_range_attribution.py::test_range_verdict_reports_a_hit_planted_in_the_scanned_range
1 failed in 36.36s
--- restoring ---
RESTORE OK

excerpt - 1,726 characters of output in total

• uv run pytest -q -n 4 tests/test_history_gate_range_attribution.py tests/test_history_gate_hit_report.py 2>&1 | tail -20

warning: `VIRTUAL_ENV=/Users/eyalgolan/git/<redacted>-public/.venv` does not match the project environment path `.venv` and will be ignored; use `--active` to target the active environment instead
bringing up nodes...
bringing up nodes...

.................................                                        [100%]
33 passed in 13.94s

Not verified: everything below is a limit of this section, listed whether or not it bit this attempt.

• no command recognised as e2e, http, typecheck, lint, build was recorded - and a recorded command is shown with its middle omitted, so a check inside the omitted part cannot be ruled out
• an entry shows that a command LINE was submitted to the shell and what came back - never that the check recognised inside it RAN, and never that it was the RIGHT command: pytest -k test_nothing prints a clean run, and a recorded command line may name a check the shell never reached yet is still counted - TEN SHAPES WERE DRIVEN against bash 3.2.57 with the check replaced by a marker-printing stub and the marker was absent in every one: a failed &&, a taken ||, an exit, an exec, an exit inside a sourced script, a syntax error that aborts the REST of the line, a multi-line if false, a case that matches nothing, set -e aborting an earlier command, and set -u on an unset variable; that list is MEASURED, NOT EXHAUSTIVE, because this module is not bash, so a kind this section does NOT list as missing is a kind some recorded line named, which is not the same as a kind that ran
• the text is the coder's: the session chose the command string and, through echo/printf, can choose the output too. Both are shown as inert text, and no entry ASSERTS a pass, a fail, or an exit status - pytest -q | tail -3 exits with tail's status, Error: Exit code 1 is a line IN THE OUTPUT, and where the harness reported a timeout or an interruption instead of output that report is appended to the captured text in square brackets. Read the output
• recognition reads the command line ONLY - it never looks inside what a command runs, so bash -c 'uv run pytest -q' leaves no receipt at all while make test leaves one that names make and not the recipe it ran; and the other way, a check merely NAMED in a heredoc body, or in a quoted string that happens to spell a shell separator, can be recorded as though it ran
• commands run inside a spawned subagent are deliberately excluded, so delegated work leaves no receipt here; a command the harness refused to run (blocked, or permission denied) leaves none, because it never ran; and only a command the HARNESS backgrounded leaves no receipt at all - it hands back a task id instead of output. A trailing & YOU wrote is NOT that and is NOT excluded: pytest -q & is recorded and headed test, and may still have been running when the harness returned
• the COMMAND and the output are both redacted and bounded before they are stored - an excerpt is not the full log, a credential-shaped string may have been masked out of either, a command over 400 characters is shortened in the middle, each command is displayed on ONE line with its newlines folded to spaces (so it may not re-run as written), and invisible and direction-changing characters are stripped before display; look-alike letters are NOT detected
• nothing here checks that these commands exercise the diff - no receipt is compared against the files this PR changes; no interactive UI check was performed (no_human never drives a browser at your change except testing/ui_evidence.py's walk, reported as its own evidence, not a receipt; the only other page it drives is a CI server's login form, and the board it opens without driving, so an e2e entry is the project's harness printing its result, not a human-style walkthrough); and no_human's own test run, CI, and the independent review are separate signals - this section covers only the coder session's own commands
• at most 200 receipts are recorded per attempt; past that the observer stops recording, and this section says so above when the limit was reached

See the PR body's Evidence table for the orchestrator's own test run.

Independent review — PASSED (1 round) on 4c2bdd7

A different model, fresh context, commit, push and merge refused at the tool call, told to refute "done". This is the checklist the gate decided on; no_human never merges — a human does.