Reconstruction pass: make the gates tell the truth, make the package self-contained, keep the one-hop symptom index

An external audit of this repository landed on four structural findings, and this pass is the
response to them. Everything below was re-measured on this tree before being acted on; two of the
audit's claims needed a correction of my own, and both corrections are noted in place.

**The gates were optimistic, and one of them did not read content at all.**

  doctor.py    the capability table was hard-coded (`repack + sign = bool(java) and python`), so a
               machine with a JDK and no zipalign/apksigner was told it could re-sign; the script
               inventory was a hand-written list of 28 against 51 files on disk, omitting the nine
               most important scripts. It now scans the directory (`registered/checked = 52 / 52`,
               unjudgeable entries counted as `unknown`, never skipped) and resolves capability as a
               real requires-closure, reporting BLOCKED with the missing piece and an installable
               next step. Measured: the old doctor reports `[OK] repack + sign` on a JDK-only PATH
               where keytool, jarsigner, zipalign, apksigner and all eight smali jars are absent.
               The report is allowed to look bad; that is the point.
  check_commands.py (new)
               anchor checks prove a section exists, never that a command runs. This parses each
               script's argparse table with `ast` and checks every documented invocation against it.
               146 commands / 41 interfaces, 0 drift. It caught `sig_probe.py --live --pkg`, which
               does not exist -- the reference had been wrong since it was written.
  check_routing.py (new)
               the symptom index now has one home (`SKILL.md`, because symptom -> file must stay one
               hop) and the inventory has another (`references/routing.md`). This gate proves the
               mirror agrees, that every reference file is named, and that every script is named.

**so_constpatch.py was corrupting APKs it rebuilt.** The writer copied filename and timestamp onto a
fresh ZipInfo and recompressed the whole archive: 435-entry real package, `extra` lost on 205 entries,
`external_attr` rewritten on all 435, 132 untouched entries recompressed, and `resources.arsc` pushed
to an unaligned offset. Android's own `zipalign -c -v 4` fails the old output and passes the new one;
a 26-case fixture suite pins it, and `extractNativeLibs="false"` is now refused by default with a
pointer at `repack.py`. Fixing it exposed the same defect in `repack.py`, where `^lib/[^/]+\.so$`
cannot match `lib/arm64-v8a/*.so` -- so its alignment handling had never once fired on a real APK.

**The package was not self-contained.** 59 references pointed at the repository root, which `npx
skills add` does not install. They now converge on `references/evidence-summary.md` (capability,
conclusion, strength, evidence that ships) plus `evidence/` with the machine-readable companions: 8
references remain and each is explicitly labelled "repository root, not shipped". A drill in which
the new files were copied to an isolated directory and asked three questions answered all three.

**SKILL.md: 572 lines / 15,733 tokens -> 398 lines / ~10,900 tokens**, without losing the two things
that had to stay. The symptom index remains inline; the reference and script indexes moved to the
on-demand routing file. The freed budget went into making the rules *stronger*, not shorter: the
constraints section is now a table of thirteen non-negotiable constraints each paired with the check
that catches a violation, and the tooling obligations are instructions whose violation is a defect
(hedging language in SKILL.md: 1 line, down from a paragraph of hedged advice). Long references gained
a navigation head so a reader can decide from the top whether to load one.

**New: an executable test layer.** `tests/` (unit, CLI contract, no-device integration) plus
`pytest.ini`, `tests/run_tests.py`, a minimal CI job, and `evals/evals.json` written but never run --
stated as such in the file. 608 passed / 158 skipped / 25 xfailed. The 25 xfails are a work order, not
a pass: thirteen scripts report a missing input by exiting 1 instead of 2, seven let a traceback
escape, four treat `--help` as an error, and `preflight --json` lacks a `status` field. A regression
row for `dex_patch_bytes.py` reproduces a previous artifact byte-for-byte
(`67ae1e58e7a44777e5ac8f9f823ab21e…`) and additionally pins the four bytes that may move.

**Two real defects the new tests found in the tooling itself.** `check_repo.py` decoded subprocess
output with the locale encoding, so on a GBK console a script printing an em dash in its usage line
was reported as crashed; it now decodes UTF-8 explicitly. And a CLI contract test drove a *real
attached device* through `coldstart.py` -- logcat, screenshots, five orphan `adb logcat` processes --
which is fixed twice over: child processes run in a private temporary cwd, and the fifteen
device-bound scripts are never executed beyond `--help`/argument parsing.

**Quoting on the device side is now one implementation.** `devsh.su` escaped only double quotes, so
`;`, `$( )` and backticks in a caller-supplied value still reached the device shell while
`coldstart.Device.shell` quoted correctly. Both use `scripts/device_shell.py` now: POSIX single-quote
escaping, argv arrays by default, and identifier validation that refuses a malformed package, path or
component with exit 2 instead of a traceback. The integration test that used to pin the *defect* (it
asserted `su -c "…"` with `\"` escapes) now asserts the property, across five metacharacter payloads.

Gates: check_repo 0 problems (including the content-reading identity scan over git ls-files),
check_refs 0 dangling / 0 warning, check_routing coherent (53 symptom / 45 reference / 54 script rows),
check_budget 0 failure / 0 warning with the always-loaded body now measured in tokens rather than in
"narrative lines with the index excluded", check_commands 146 checked / 0 drift, pytest 608 passed.

Not done, and stated rather than implied: CI has never run on a runner; the evals have never been
executed; the 25 xfails are open; no device-side install/launch was performed in this pass.

.github/workflows/ci.yml

@@ -0,0 +1,86 @@
+# CI for the apk-reverse skill repository.
+#
+# WHAT THIS RUNS
+#   * the repository's own structural gates: check_repo.py, check_refs.py,
+#     check_budget.py, check_routing.py, and check_commands.py when it is present
+#   * the pytest suite: tests/unit, tests/cli, tests/integration
+#
+# WHAT THIS DOES NOT RUN, ON PURPOSE
+#   * anything needing a device, adb, frida, a rooted phone, a network peer or a
+#     real sample APK. Those routes are recorded experiments in tests/benchmark.md
+#     and their status is documented in
+#     docs/tool-verification/EXTENSION-test-harness.md.
+#   * anything about agent behaviour. Whether a model routes, classifies or
+#     refuses correctly is not observable from a unit test; skills/apk-reverse/
+#     evals/evals.json is the skeleton for that, and it has never been executed.
+#   * the dex_patch_bytes regression hash. Its input lives in git-ignored
+#     tools/_work/bench/, so that test self-skips here with a printed reason
+#     rather than failing. Run it locally after downloading the sample.
+#
+# Coverage is therefore: static structure, documentation anchors, budget, command
+# references, pure parsers, CLI contracts and no-device flows. Nothing more is
+# claimed.
+name: ci
+
+on:
+  push:
+  pull_request:
+
+permissions:
+  contents: read
+
+jobs:
+  test:
+    name: python ${{ matrix.python-version }}
+    runs-on: ubuntu-latest
+    strategy:
+      fail-fast: false
+      matrix:
+        python-version: ["3.11", "3.13"]
+
+    env:
+      PYTHONIOENCODING: utf-8
+      PYTHONDONTWRITEBYTECODE: "1"
+      # The gates spawn sibling scripts and decode their output. On a Windows
+      # runner the default codec is the console's (GBK here), and one script's
+      # --help text is not GBK-decodable, which is reported as a crash in that
+      # script rather than as a harness decoding problem. Pinning UTF-8 removes
+      # the whole class of failure on every platform.
+      PYTHONUTF8: "1"
+
+    steps:
+      - uses: actions/checkout@v4
+
+      - uses: actions/setup-python@v5
+        with:
+          python-version: ${{ matrix.python-version }}
+
+      - name: Install test dependencies
+        run: python -m pip install --upgrade pip pytest
+
+      - name: Structural gate - repository layout
+        run: python check_repo.py
+
+      - name: Structural gate - documentation anchors
+        run: python check_refs.py
+
+      - name: Structural gate - context budget
+        run: python check_budget.py
+
+      - name: Structural gate - documented commands (if present)
+        run: |
+          if [ -f check_commands.py ]; then
+            python check_commands.py
+          else
+            echo "check_commands.py is not present in this revision; skipping."
+          fi
+
+      - name: Structural gate - routing inventory matches the entry point
+        run: python check_routing.py
+
+      - name: Regression suite
+        run: python -m pytest -q
+
+      - name: Summarise skipped and expected-failure coverage
+        if: always()
+        run: python -m pytest -q -ra --no-header -p no:cacheprovider 2>&1 | tail -n 40