rtk 0.51.0: make the grep file-list exactness control lossless (follow-up to #693) - #698

PR #693 / native-token-tools run 37226622626 failed because our control
required unchanged grep file-list stdout after upstream deliberately folded
the shared directory prefix. Keep coordinator option (a) and the 0.51.0 pin.

Parse the documented <prefix> (N files) header, expand each tail by literal
prefix concatenation, and compare the complete reconstructed native path set.
Check successful exits, distinct native paths, and the header/tail counts.
Use LF records to preserve CR corruption. Keep both diff predicates unchanged
and retain exact native/proxy output and exit recovery for every case.
Add a native three-file guard fixture whose first filename ends in CR; require
the CR signal and byte-identical verbatim output, including the proxy arm.

Pinned original sources:
- rtk-ai/rtk v0.51.0, e001f773f80b22b7dc4c7a79521b30e35aaef026;
  published 2026-10-02T13:33:54Z, verified through the GitHub release/tag API.
- feac25d15f254dcdbb6c6629b91328a821eb24ce:
  src/cmds/system/search.rs:612-674 and :1345-1354;
  tests/search_file_list_fold_test.rs:43-66 at v0.51.0.
- 3223a80145e482762abbf50ac01656360f906f14:
  src/cmds/system/search.rs:631-639 and :1337-1342 at v0.51.0.

Append a dated limitation to the existing 0.51.0 receipt and add the dated
decision note. Historical receipt claims and observations remain intact.

Returned evidence:
- Old assertion with the 0.51.0 synthetic receipt shape: one test fails,
  exit 1 (.bounded-job-035/tmp/red-unittest.txt).
- Old native exactness fixture reproduces native exit 0 / 60 bytes / 3 lines,
  RTK exit 0 / 45 bytes / 4 lines, and byte-identical proxy output; exit 1
  at the original control (.bounded-job-035/tmp/red-native.txt and
  .bounded-job-035/tmp/native-red/results/receipt.json).
- Guard test before its fixture/check addition fails, exit 1; an injected
  folded-tail CR also fails before replacing splitlines() with LF splitting
  (red-guard-unittest.txt and red-fold-cr-unittest.txt under the job TMPDIR).
- Final native exactness fixture: nine checks pass, 11 cases / 33 arms /
  73 commands, exit 0 (.bounded-job-035/tmp/green-native-final.txt and
  .bounded-job-035/tmp/native-green-final/results/receipt.json).
  Independent raw stdout hashes confirm exact proxy recovery and CR refusal;
  diff output hashes and exits match before and after the repair.
- python3 -m unittest tests.test_native_token_ci: 49 tests pass, exit 0.
- All four referencing modules were run: 914 tests, 198 failures, 35 skips,
  exit 1. Every failure is in tests.test_token_e2e_grader and refuses private
  fixture output with E_PATH reason=work_tree. An existing /tmp/.git entry
  triggers tools/token-e2e/frozen_checks.py:1311 despite native git -C /tmp
  rev-parse --is-inside-work-tree returning exit 128. An isolated failing
  case reproduces the refusal (grader-isolated-failure.txt). Git diff confirms
  that the grader, its tests and the privacy guard are unchanged from HEAD.
- The other three referencing modules pass independently: python3 -m unittest
  tests.test_adoption_bootstrap_macos tests.test_agentsview_qualification
  tests.test_workflow_hardening: 290 tests, 35 skipped, exit 0.
- python3 scripts/validate.py: exit 1, registry drift only (seven findings):
  SHA-256 and byte counts for the changed script, tests and qualification
  receipt, plus the appended receipt-limitations mirror. Refresh these in
  manifests/evidence.json at integration; that file remains untouched here.
- git diff --check: exit 0.

Final acceptance logs: acceptance-native-token-ci-final.txt,
acceptance-referencing-modules-final.txt,
acceptance-other-referencing-modules.txt and acceptance-validate-final.txt
under .bounded-job-035/tmp. Earlier failures and reruns remain alongside them.

The requested TMPDIR path is a symlink to isolated /tmp storage, and acceptance
uses its canonical path. Native npm uses a job-local cache and distinct empty
user/global config files. The initial in-checkout TMPDIR attempt failed the
grader preflight and native npm packing; a /dev/null-for-both-configs retry
failed with npm's double-loading rejection. Generated job artifacts are ignored
for publication. The shared /tmp/.git and all refusal guards remain untouched.

Workflow: bounded search-first / diagnosing-bugs / tdd at the user-selected
exactness seam, based on original pinned source. Scoped ai-memory retrieval was
unavailable under the tool approval policy; no configuration was changed.
Provider usage is unknown. The completeness critic in the dated note keeps
native rg/-L/NUL/OSC 8, other hosts and model interpretation outside this round.

Residual: agent-facing expansion of folded paths is unmeasured. An E2E
observation of a wrong expanded path leads to adding grep -l/-rl to the hook
exclusions. No upstream Cargo suite or native-model acceptance is claimed.

No commit was created. Branch c5/token-pin-move-20261004 remains at
bbb299d. Changes are left in the working tree.

Co-Authored-By: Claude Opus 5.5 <[email protected]>

…facts; describe only main's history in the guide

The Claude delta read found 421e4ee and 45a7218 cited as sources although the rebase onto 38ac9ac replaced both
(421e4ee is 5a4add3 with the same patch; 5e6afa4 replaced 45a7218). Each citation now says pre-rebase, the guide
describes #693 then #698, and the note records validate passing after the registry commit.

Co-Authored-By: Claude Opus 5.5 <[email protected]>

…d artifact README

Co-Authored-By: Claude Opus 5.5 <[email protected]>

…egistry plus the owned rows)

Co-Authored-By: Claude Opus 5.5 <[email protected]>