plan: change: a fourth shape of the defect, and a build recipe that works

Building #1155's commit 11 found what three static passes had not: its
headers include TcpClassicAlgorithmBase.h, TcpAlgorithmBase.h and
TcpCubic.h before those files exist. Phase B was written against the
names the rename would later give the files, so commits 11 to 14
include headers from their own future. check-includes.sh now finds it
without a compiler, and step 1h3 carries the repair, which is small:
point the includes at the names that exist there and let the rename
commit update them, which is what the rename commit is for.

The gate also reports the repair's own move commit and that one is
right: a move that changes no content cannot fix its own includes, and
PR-SERIES-BUILDS exempts it.

The build recipe is written down because three traps cost an hour.
Build from the repository root, never from src/, because the root
makefile generates src/inet/features.h and the sub-make does not.
Objects land in src/out rather than out. A libINET.so left in src/ by
an earlier attempt makes the whole tree look current, so make exits 0
having done nothing. And git checkout leaves an unchanged file's mtime
alone, so after moving to an older commit the differing files must be
touched before make will rebuild them.

With that, the head builds -- 1693 sources, no errors -- and per-commit
builds work, which is what unblocked 1h3 and what step 1i needs.

Plan: plan/pending/pr-1155-resolve-audit-findings.md
Change: plan | behavior.change | - | pr-1155-findings

Author: levy

plan/pending/pr-1155-resolve-audit-findings.md

@@ -228,6 +228,22 @@ The same tree-by-tree rebuild moved each field's declaration to the commit that
 under a `// declared here for the code above that already uses it:` header. **The final tree is
 byte-identical to the audited head** — the repair adds nothing at the end, only earlier.
 
+**1h3. Repair the four forward includes (F-9). — open, and small.**
+
+Commits 11 to 14 include `TcpClassicAlgorithmBase.h`, `TcpAlgorithmBase.h` and `TcpCubic.h`
+before those files exist, which
+[check-includes.sh](../../doc/project/enforcement/check-includes.sh) now reports. Phase B was
+written against the names the rename would later give the files.
+
+Two ways to fix it, and the second is better. Point the includes at the names that exist at those
+commits and let the rename commit update them — which is what the rename commit is for, and it
+already touches those files' neighbours. The other way, moving the rename earlier, reorders the
+series for a smaller reason than it deserves.
+
+The gate also reports the repair's own move commit, commit 15, and that one is correct: a move
+that changes no content cannot fix its own includes, and
+[PR-SERIES-BUILDS](../../doc/project/rule/pull-request.md#pr-series-builds) exempts it.
+
 **1i. Re-cut phases B and C by feature (F-10). — open, and it is the large one.**
 
 The parameter half of the partition is repaired; the logic half is not. Commit 18 still carries
@@ -440,10 +456,26 @@ at the commit that uses it, across all 63 commits, by
 [check-ned-params.sh](../../doc/project/enforcement/check-ned-params.sh) and a matching pass over
 the state messages.
 
-**Setting up a reliable per-commit build is now the blocker for step 1i**, and
-[TR-CI-EVERY-COMMIT](../../doc/project/rule/testing.md#tr-ci-every-commit) wants it regardless. The
-recipe needs three things this session did not get right: build in `src/`, never leave a library
-in `src/`, and regenerate the `_m` files whenever the `.msg` files change under it.
+**The build recipe, which now works.** From the repository **root**, never from `src/`, because
+the root makefile generates `src/inet/features.h` and the sub-make does not:
+
+```bash
+export PATH=/home/levy/workspace/omnetpp/bin:$PATH OMNETPP_ROOT=/home/levy/workspace/omnetpp
+export LD_LIBRARY_PATH=/home/levy/workspace/omnetpp/lib:$LD_LIBRARY_PATH
+make MODE=release -j32
+```
+
+Three traps cost this session an hour. Objects go to **`src/out`**, not `out`. A `libINET.so`
+left in `src/` by an earlier attempt makes the whole tree look current, so `make` exits 0 having
+done nothing. And `git checkout` leaves the mtime of an unchanged file alone, so after moving to
+an older commit the files that differ must be touched before `make` will rebuild them:
+
+```bash
+git diff --name-only HEAD <newer-ref> -- 'src/*' | xargs -r touch
+```
+
+With that, the head builds — **1693 sources, 0 errors** — and a per-commit build works, which is
+what unblocked 1h3 above.
 
 **Done when** every commit that moves a recorded expectation carries it with a row-level
 explanation, `check-commits.sh` reports no `PR-SPLIT-BASELINE` violation, and