diff --git a/perforce/LEARNINGS.md b/perforce/LEARNINGS.md
index f1d5e95..57f4b00 100644
--- a/perforce/LEARNINGS.md
+++ b/perforce/LEARNINGS.md
@@ -12,6 +12,53 @@ them.
 
 ---
 
+## 2026-09-14 - the typemap is one server-wide table, and a p4d container image is a second writer `[integration]`
+
+Learned running a Git->Perforce mirror against a p4d that ships as a Docker image and
+re-applies its own engine typemap preset on every container start, not just first
+provision.
+
+- **`p4 typemap` is a single server-wide table, so "add my depot's rows" is a
+  read-modify-write against state someone else also owns.** The documented pattern
+  (`p4 typemap -o | <insert rows> | p4 typemap -i`) has no compare-and-swap - any other
+  writer that reads, edits, and writes the whole table drops rows it doesn't know about.
+- **The other writer here was the p4d image itself.** Its startup script re-applies an
+  engine preset on every boot, and a restart taken just to pick up an unrelated image
+  upgrade silently removed six depot-scoped rows a different provisioning script had
+  appended - nothing failed at the time, the rows were simply gone.
+- **The exposure window measured ~8 minutes, not seconds**, timestamped on one real
+  restart. Because typemap only applies to newly-added files, a submit landing inside the
+  window leaves mis-typed files behind after the window closes, needing `p4 edit -t
+  <type>` to fix.
+- **So verify the rows before the first `add` and fail the run - don't warn and
+  continue.** A sync job that warns will add a whole tree with the wrong types during a
+  restart window; a hard fail costs one retried run instead.
+- **Without a spec depot, there's no history or author for typemap/protections edits** -
+  `p4 typemap -o` shows current state only, and forensics stop dead. `p4 depots` tells
+  you in one command whether you have one.
+
+## 2026-09-14 - "Wrong number of words", and what p4's exit codes do under `set -e` `[skill]`
+
+Three diagnostic details that cost real time chasing the entry above.
+
+- **"Wrong number of words for field 'X'" doesn't always mean what `p4-observe`'s
+  trigger-table advice says.** That advice is right for the *trigger* table (an
+  unquoted command containing spaces). For the **typemap** spec, the identical string
+  *is* whitespace: rows must be TAB-indented. Read the error against the spec you're
+  actually editing.
+- **`p4` exits 1 with empty stdout on both "cannot connect" and "no ticket,"** verified
+  against a live server. That matters under `set -euo pipefail`: an assignment like
+  `current="$(p4 typemap -o)"` **aborts** on those failures rather than yielding an
+  empty string - so a script that reaches its "rows are missing" branch really did read
+  successfully; it's not a swallowed connection error.
+- **Idempotency keyed on a description trailer must accept every width its writers
+  actually produce.** A mirror recording `Original-Commit: <sha>` and matching it back
+  with a 40-char-only pattern missed a code path that wrote an abbreviated sha, read
+  that as "nothing synced," and replayed the branch from its root commit - which then
+  failed inside git-p4's apply step with "No valid patches in input." Matching a
+  variable-width sha and resolving it to the full value locally fixed it; treat an
+  unresolvable trailer as a hard error, not as "nothing synced."
+
 ## 2026-07-16 - p4 trigger/webhook ingestion `[integration]`
 
 Learned wiring Perforce change-triggers into a webhook ingestion pipeline:
diff --git a/unreal/LEARNINGS.md b/unreal/LEARNINGS.md
index 9fa638c..2cc9a65 100644
--- a/unreal/LEARNINGS.md
+++ b/unreal/LEARNINGS.md
@@ -170,3 +170,10 @@ No secrets - reference credential *locations*, never paste them.
       is a node action; bare `nvidia/cuda` images die on step one). catthehacker
       act-22.04 + the NVIDIA container toolkit + `container.options: "--gpus all"` works;
       driver injection gives nvidia-smi-level access, CUDA only if the image adds it.
+15. **Two traps when authoring the automation scripts themselves, both of which hang
+    instead of erroring.** A PowerShell `[Parameter(Mandatory=$true)]` argument
+    silently prompts on stdin when omitted, so a non-interactive SSH invocation just
+    blocks forever - validate the argument by hand rather than trusting the parameter
+    binder to catch a missing value. And `ssh -n` redirects stdin from `/dev/null`,
+    which silently swallows a heredoc meant to feed a remote `bash -s` - the remote
+    script runs with no input and nothing reports an error.