[Nicholas023]

What was wrong

PR #907 made a PTC scope with dynamic loading off publish the configured allowedTools pool as resident. That answer to #904 is being withdrawn: the behaviour is not the one we want on main.

The change

git revert -m 1 d670333b — one commit, eight files, the exact inverse of the merge.

Proof of the inverse: git diff --stat d670333b^1 8818e09c prints nothing, so every reverted file is byte-identical to its state before #907.

• src/publication/ptc.ts no longer contains the exposed computation.
• test/modes.test.mjs is back to asserting that a loader-less scope publishes no selectable pool.
• The four documents, the two locales and the CHANGELOG entry return to their pre-fix(ptc): publish the allowed pool as resident when dynamic loading is off #907 text.

Deliberately not done here: no replacement behaviour is added. This PR only removes #907, so the follow-up can be designed on its own terms.

Validation

The revert is a pure inverse, so the tree this PR produces is the exact tree CI accepted before #907:

$ git diff --stat d670333b^1 8818e09c
$                              # empty: no difference from the pre-merge tree

$ grep -n "exposed" src/publication/ptc.ts
$                              # empty: the resident-pool computation is gone

$ grep -n "loader-less scope publishes" test/modes.test.mjs
396:    assert.equal(names(agent).includes(publicToolName('write_file')), false, 'a loader-less scope publishes no selectable pool')

Affected suites, run locally against the reverted tree:

$ node test/run.mjs modes
dsh-ops modes: all 12 checks passed
dsh-ops release-tar-modes: all 4 checks passed

$ node test/run.mjs resident-switch
dsh-ops test: 1 suite(s) passed

$ node test/run.mjs ptc-declaration-surface
dsh-ops test: 1 suite(s) passed

$ node test/run.mjs bench-tools
dsh-ops bench-tools: all 12 checks passed   # 6.47s, exit 0

CI — every job on this PR:

CI required          pass   3s
node-floor           pass   3m21s
rust-lint            pass   28s
source-size          pass   6s
tools-trial          pass   47s
validate (darwin-arm64)  pass   4m1s
validate (darwin-x64)    pass   9m0s
validate (linux-x64)     pass   3m25s
validate (win32-x64)     pass   8m4s
workflow-lint        pass   10s

The first validate (darwin-arm64) attempt failed on an exit hang, not an assertion: bench-tools.test.mjs reported all 12 checks passed and then kept its process alive until the 300 s suite deadline stopped it. The rerun above passed in 4m1s. Nothing in this diff touches test/ or bench/, and main at the identical commit d670333b passed the same job in 3m30s, so it is pre-existing harness flakiness rather than a regression from this revert.

Environment

• OS / 架构：Windows x64
• Node：v24.21.0
• 基于上游 main 的 commit：d670333b
• 安装通道：npm ci --ignore-scripts（本地临时工作树）

---