Add observation-window hygiene and non-Java update gates

references/long-task-discipline.md
  New section: "Keep the observation window clean". Every conclusion rests on "the
  build I am looking at is the build I made", and that assumption fails silently:
  something else installs/reverts/clears the target mid-capture, the process was never
  actually restarted, or the installed artifact is not the one on disk. Guards, cheapest
  first: pin identity by hash (not filename, and immediately before the window, not
  hours earlier); declare a capture window and run no install/uninstall/clear inside it;
  timestamp it; prefer one continuous capture over several short ones. When a window
  turns out to be contaminated, discard and re-capture rather than salvage - a re-run
  costs minutes, a wrong conclusion costs the rest of the task and gets re-derived from
  the same bad evidence.

references/updates-and-forced-upgrade.md
  New Step 6: "when the check is not where you expect", covering three shapes that break
  the flip-the-comparison plan.
  - The check is native: a declared update entry point with zero Java callers means the
    invocation lives in a .so. Ordered remedies, starting with the one that needs no
    native work at all - attack the response, or feed the check a version it likes. "No
    Java caller" is not "unpatchable".
  - The dialog hands off to a browser or WebView: there is no installer call to find,
    and logcat truncates the URL to `dat=https://host/...`, so a grep gives you a host
    and no path. Approach table for recovering the real URL, with hooking intent
    construction as the reliable one.
  - Server-issued gate with a local fallback: read the failure branch first. If failure
    defaults to "no update", failing the request is the whole fix and also gives you an
    offline verification signal; if it defaults to "must update", it makes things worse.

references/long-task-discipline.md

@@ -193,6 +193,39 @@ Three rules that follow:
 - **Prefer a bounded observable over a fixed sleep.** Poll the state you actually care about, with a
   cap — and prefer a sample you can inspect over a duration you hope is right.
 
+## Keep the observation window clean
+
+Every conclusion in this skill rests on "the build I am looking at is the build I made". That
+assumption is silently false more often than any other, and a contaminated window does not look
+contaminated — it looks like a result.
+
+Three ways it happens, all of which have produced confidently wrong findings:
+
+- **Something else touched the target while you were observing.** Another process, another agent,
+  another window of your own work installed, reverted, restored or cleared the app. Your screenshots
+  and logs are then a mixture of two states and describe neither.
+- **You are looking at a stale process.** The app was never actually restarted, so the "before" and
+  "after" captures come from the same run.
+- **The artifact on the device is not the artifact on disk.** An install that reported success, or was
+  skipped because the version matched, leaves the previous build running.
+
+Guards, in order of cost:
+
+1. **Pin the identity, not the filename.** Hash the artifact you built, hash the artifact you intend
+   to test, and hash what is actually installed (pull the installed APK or read its digest) — the
+   filename proves nothing. Do this immediately before the observation window, not hours earlier.
+2. **Take the window deliberately.** Before a capture sequence that a conclusion will rest on, state
+   (to yourself or in the record) that the next N seconds are for this observation only, and do not
+   run another install/uninstall/clear inside it.
+3. **Timestamp the window.** Record the wall-clock start and end. It costs one line and it is the only
+   way to later notice that a teammate, a background job or your own earlier command landed inside it.
+4. **Prefer one continuous capture over several short ones.** A restarted capture invites a restarted
+   state; a single sequence cannot straddle an install it did not perform.
+
+When you discover a window was contaminated, **the correct action is to discard it and re-capture**,
+not to salvage it. A re-run costs minutes; a wrong conclusion costs the rest of the task, and it will
+be re-derived from the same bad evidence because the record says it was observed.
+
 ## Handover
 
 A handover is the record plus three things, written for someone with **no** memory of the task:

references/updates-and-forced-upgrade.md

@@ -165,6 +165,69 @@ Update suppression changes behaviour a user can observe ("it never tells me abou
 in the delivery notes, in one line, together with anything else you neutralised. If the app also has a
 legitimate reason to update (security fixes), say that the build will not pick them up automatically.
 
+## Step 6: when the check is not where you expect
+
+Three shapes that break the "find the Java comparison and flip it" plan. Recognise them early,
+because each has a different cheapest remedy.
+
+### The check is native
+
+Symptom: you grep the whole dex tree for the version-comparison routine, the update endpoint, or the
+dialog text and find **the method declared, and zero callers**. A `static native` update entry point
+with no Java caller means the invocation lives in a `.so` (`registerNatives` from a static block is
+the usual giveaway — see `code-virtualization-and-custom-linkers.md`).
+
+Options, cheapest first:
+
+1. **Attack the response, not the check.** Block or rewrite the update endpoint at the network layer,
+   or empty the version/config response. The native code still runs and still decides "no update".
+   This needs no native work at all and survives a rebuild — check this before reaching for a
+   disassembler.
+2. **Send the check a version it likes.** If it compares against a value the app itself supplies
+   (a config key, a stored preference, a build field), setting that value is a one-line change.
+3. **Patch the native comparison.** Last, because it is the most expensive and the most fragile.
+
+Do not conclude "unpatchable" from "no Java caller". The gate usually reads a value you *can* control.
+
+### The update hands off to a browser or a WebView
+
+Very common in this category: the dialog's button fires an `ACTION_VIEW` intent at a download URL,
+or loads it in an in-app WebView. Two consequences:
+
+- **You will not find an installer call** in the app, so searching for `PackageInstaller`,
+  `REQUEST_INSTALL_PACKAGES` or a download service finds nothing. That is not evidence the dialog is
+  harmless; the browser does the install.
+- **Logcat truncates the URL.** `ActivityTaskManager` prints `dat=https://host/...` and elides the
+  rest, so a naive grep gives you a host and no path. Do not build a plan on a truncated URL.
+
+Ways to get the real URL, cheapest first:
+
+| Approach | Notes |
+|---|---|
+| Hook `Intent` construction / `Context.startActivity` and log `intent.getDataString()` | Most reliable. A truncated log line is a *logging* limit, not a limit on what you can read |
+| `dumpsys activity activities` / recents while the browser is still in the stack | Free; works only if the activity is still alive |
+| The browser's own history database (root) | Survives the browser being closed; needs the right DB, and modern browsers may encrypt or prune it |
+| Read the update *response* instead | Often easier: the URL came from a JSON field you can fetch or hook directly |
+
+**Deliberately tapping the button is a legitimate probe** — but treat it as a measurement, not a
+side effect: capture the intent (or the screen) in the same window, and know that on some devices a
+second tap resumes an already-started download rather than re-issuing the intent.
+
+### The gate is server-issued with a local fallback
+
+`fetchEnabled()`-style helpers frequently **default to a local value on any failure** (non-200,
+timeout, malformed body, signature mismatch). Two things follow:
+
+- **Read the failure branch before you touch anything.** If failure ⇒ "no update" or "no ad", then
+  merely making the request fail is a complete fix, and it is far cheaper and more robust than
+  patching the comparison.
+- **The fallback also tells you how to test offline.** Disconnect the device and relaunch: if the
+  gate disappears, the fallback is benign and you have a zero-code workaround as well as a
+  verification signal.
+
+Conversely, if failure ⇒ "must update", failing the request makes things worse, and you must patch the
+decision instead.
+
 ## Checklist
 
 - [ ] Update vocabulary searched in **both** string encodings, in the right layer for the runtime