mechanical-compiler/docs/HANDOFF.md

HANDOFF

This document is rewritten in place each session. It is state, not a log.

It is the only handoff you need to read. Dated handoffs in docs/archive/ are historical and are not required reading — do not diff them against this to work out what is true. If something here is wrong, correct it here.

Last updated 2026-08-20 at commit 114189c.

---

1. Invocation — read before typing anything

These are facts, not examples. Getting one wrong produces an error that looks like a problem with the repository or the containers.

mechcomp has nologin as its shell. Use runuser -u mechcomp -- <cmd>, which executes directly. su - mechcomp cannot work and fails with This account is currently not available — a message that looks like a broken account and is not. See F-035.

All repository operations run as mechcomp, never root. The clone is at /var/www/mechcomp in CT 100. Never add a git safe.directory exception to work around an ownership complaint; fix the ownership (F-008).

mechcomp's home is the working tree. Anything writing to $HOME writes into the repository. .cache/, .local/ and .ssh/ are in .gitignore for that reason (F-029). Git identity is set --local for the same reason — --global would write into the tree.

verify.sh is mode 100644. Invoke it as bash tools/reference-toolchain/verify.sh, never ./tools/....

Gitea SSH is port 42022. Remotes need ssh://git@host:42022/owner/repo.git; the git@host:path shorthand cannot carry a port. CT 100 pushes with deploy key srv-b-ct100.

Deploy tokens in Gitea are account-level, under user Settings. Repository settings offer deploy keys only.

Files reach CT 100 by upload, then pct push, then chown. pct push writes as root, so pct exec 100 -- chown mechcomp:mechcomp <path> immediately after, every time.

Every network command needs an explicit timeout (F-030). One without hung the operator's shell.

Logs are in journalctl, not /var/log/. Proxmox ships without rsyslog (F-024).

Long-running jobs go to systemd-run --unit=<name> --collect, not nohup or setsid. pct exec tears those down when it exits; systemd owns the job and the output lands in the journal.

Docker runs as root in CT 100 only. mechcomp cannot reach the daemon. CT 101 has no keyctl and cannot run it at all.

Assert the guest is running before interpreting any pct exec result (F-027). A command that fails because the container is stopped otherwise reads as a pass.

Useful canonical form:

if [ "$(pct status 100 | awk '{print $2}')" != "running" ]; then
  echo "CT 100 NOT RUNNING - stop here."
else
  pct exec 100 -- runuser -u mechcomp -- git -C /var/www/mechcomp status --short
fi

---

2. The operator's constraint

CIVICVS has a shell on srv-b and a browser-based file manager. Nothing else.

No workstation git. No SSH into a container. No IDE. No scp. Files arrive by upload to /root/incoming on srv-b; every command runs in that one shell. pct exec from that shell reaches all three containers — that is the container path, and it is not a limitation.

An assistant that assumes otherwise produces instructions the operator cannot execute. This has happened repeatedly.

Deliver code by upload, not by paste. A sixty-line heredoc containing em-dashes and nested code fences was mangled by the browser terminal. Everything delivered since as a tarball — with a stated checksum, file list, and what it overwrites — has worked without exception. State what the archive contains, where it expands, and what it overwrites, every time.

When a delivery supersedes an earlier one, ship every file in the set, not just the changed one, so the resulting state is unambiguous.

---

3. Where things stand

Infrastructure — complete, do not revisit

srv-b, Proxmox VE 8.4.0, standalone.

CT	Name	Address	Role
100	mechcomp	10.20.0.10	application, worker, Docker
101	mcproxy	10.20.0.11	reverse proxy, TLS
102	kane-fabric	10.20.0.12	separate project

All on vmbr1, a portless service bridge. srv-b is router and bastion: internet → WireGuard → srv-b → containers. Containers cannot reach the home LAN and cannot send mail.

ct-baseline.sh is read-only, runs any time, exits non-zero on divergence. Installed at /usr/local/sbin/. Last run: 62 passed, 0 failed. A property it does not check is not part of the standard — that is what makes conformance terminate rather than recur.

Settled decisions. Do not reopen any of these:

• Webmin is the operator's only remote access. Questioning it wasted a session.
• Backup is deliberately postponed. Entry condition: ct-baseline.sh exits 0. Do not raise it again.
• 4x is the end of the N-strap family.
• The --full toolchain gate's Docker root-ownership hazard is understood and handled by mounting read-only. See §6.
• Accuracy criterion: 0.01 mm over the entire set. Set by CIVICVS on 20 AUG. This is the standard the software is held to, and it is met with two orders of margin on every dimensional quantity. See docs/PRECISION.md.

The port — COMPLETE

Gitea main at 114189c. CT 100 clean and matching.

Suite: 436 passed, 30 failed. The 30 failures are expected. Do not treat a red make test as a broken port — read §7 before doing anything about it.

Module	Ported from	Contents
geom/primitives.py	sb-geom.scad	Vectors, GEO and MEMBER, sleeve and cavity paths, exact polyline distance, corner-radius derivation, monotone solver
geom/rounding.py	BOSL2 92d697c2	round_corners, _circlecorner, arc, segs, deduplicate, path_merge_collinear, is_collinear
geom/region.py	BOSL2 regions	Shapely booleans, nesting-parity decomposition, area, simplicity, hull, mitred offset, cleaning
geom/join.py	sb-join.scad	Butt joints, hull caps, fillets, derived bore, ring fit, section assembly
geom/report.py	sb-report.scad	Checks-as-values, metrics, five universal checks, ProfileRejected, report formatting
geom/core.py	sb-core.scad	PROFILE record, failure representation, centred assembly
geom/arrangements.py	sb-profiles.scad	Ring, spoke and fin arrangements, all N-generic
profiles/_common.py	both generators	Assembly pipeline and the eight shared base checks
profiles/four_x.py	strap-beam-4x.scad	Five profiles, defaults, 4x base checks
profiles/three_x.py	strap-beam-3x.scad	Six profiles, defaults, 3x base checks
profiles/__init__.py	—	build(), ProfileRejected, family registry

make deps is complete in CT 100 and must not be re-run.

---

4. What to do next

Nothing is blocked. Pick up §7 first if the operator wants it closed; otherwise the port is done and the next work is new capability.

Immediate, small, and already scoped:

1. F-034 tolerance decision (§7). Evidence is gathered; the decision is CIVICVS's. Do not implement before he chooses.
2. Rewrite F-034 in FAILURES.md around the measurements below. The entry still describes it as an unquantified hazard; it is now measured.

After that, the roadmap items — read ROADMAP.md before starting any:

• STL and STEP export. No CAD kernel is installed in CT 100 (§5).
• Dihedral parameterisation. If two panels meet at 137°, none of the eleven profiles gives you a member for it. This is the highest-value gap.
• Nodes (non-prismatic) and panels (sheet). No representation exists for either.

---

5. Facts established by porting

The public entry point is mechcomp.profiles, not mechcomp.geom

conftest.py does importorskip("mechcomp.profiles") and requires a build attribute. build(family, profile, params) -> Result and ProfileRejected are exported from mechcomp.profiles. Params use the OpenSCAD parameter names unchanged.

params carries only a case's overrides. Everything else comes from the family's declared defaults, which is why those defaults live in the port rather than the test harness.

Defaults are per-family, not shared

ring_corner_radius_mm is 2.00 in 3x and 1.25 in 4x. The two generators declare their own parameter blocks and they are not identical. Do not assume a value read from one file applies to the other. 4x also declares rectangle_aspect, which 3x has no equivalent of, and its base-check list is correspondingly one entry longer.

Check order is the reporting order

The reference reports the first failing check, not an aggregate. Base checks run before profile checks, and both run before any geometry is measured — a builder that failed returns an empty profile, and centred() on an empty profile has no members to measure.

Report values are rounded to six significant figures

The oracle records what OpenSCAD's echo printed — C's %g at default precision — not full-precision geometry. echo_num() does this; echo_vec() handles vectors, which reach the oracle as strings like '[20.5209, 20.5209, 20.5209]'.

Load-bearing, not cosmetic. VOLUME_MM3 ends in _MM3, so test_oracle.py compares it at the lengths tolerance of 1e-4, not the areas tolerance of 1e-3.

Angles are degrees; the arbitrary constants are contract

OpenSCAD trigonometry is in degrees. The port keeps degrees throughout with explicit cos_d/sin_d/tan_d so every expression matches its source line. The 44 solver iterations, the 0.999 and 0.98 scale factors, the 0.05/179.95 degree cutoffs and the 1e9 sentinel are reproduced exactly. They produced the frozen values.

Environment

Shapely 2.1.2 on GEOS 3.13.1 — keep these in step, boolean results on near-degenerate geometry can shift between GEOS releases. numpy 2.4.6. Shapely is in requirements-base.txt, the 2D path's own dependency set.

No CAD kernel is installed in CT 100. cadquery, OCP and build123d are all absent, though requirements-cad.txt says it is installed by default. That is the strictest environment for developing the 2D path and will matter when STL and STEP export begins.

What is verified against the reference

Every dimensional quantity is exact to 1e-4 mm across all 123 cases, with no exceptions. ENVELOPE_X_MM, ENVELOPE_Y_MM, MIN_WALL_ACTUAL_MM, and every profile extra: AF_*, FIN_*, SPOKE_*, RING_*, T_*. Every count is exact.

All ten rejections fire correctly, including the bespoke A Frame and T paths, which were written from the join primitives rather than from a shared arrangement. That is strong evidence those two are structurally right.

Only SECTION_AREA_MM2 and its two derivatives ever disagree. See §7.

---

6. Probing the reference directly

The pinned toolchain image is in CT 100 and OpenSCAD will echo whatever you ask it. This settled F-034 in four exchanges and will settle any later disagreement the same way. Prefer it to reasoning.

Mount the tree read-only and keep the probe on a separate writable mount — Docker runs as root, and a read-write mount against a mechcomp-owned tree is the F-033 hazard. A nested single-file mount inside a read-only mount fails with EXIT=125; give the probe its own directory.

pct exec 100 -- bash -c 'mkdir -p /tmp/probe && cat > /tmp/probe/p.scad << "EOF"
include </repo/legacy/openscad/lib/sb-core.scad>
$fn = 48;
g = sb_geo(15.875, 0.508, 1, 0.25, 1.20, 1.20, 1.20, 1.20);
echo(str("PROBE=", sb_area([sb_sleeve_path(sb_member(0,0,0), g)])));
EOF
docker run --rm -v /var/www/mechcomp:/repo:ro -v /tmp/probe:/probe -w /probe \
  mechcomp/reference-toolchain:8.0.0 \
  openscad -o /tmp/o.stl --export-format=asciistl p.scad 2>&1 | grep -E "PROBE|ERROR"'

EXIT=1 is normal — STL export fails on 2D geometry, the echoes still arrive.

The oracle itself is stable. The gate was run on 19 AUG: all 123 cases regenerate byte-identically inside the pinned image, 113 accepted and 10 rejected, the only diff being frozen and the hash containing it. verify.sh --full's restore path was broken, is now fixed, and is still unexercised (F-033).

Instrumenting the port is equally cheap. Monkeypatching mechcomp.geom.rounding._circlecorner to print its half-angle and segment count turned the F-034 diagnosis from inference into measurement in one command. The module calls it through the module global, so patching the attribute works.

---

7. F-034 — measured, and the decision is ready

The port is exact. The reference is noisy. That is the opposite of what the entry previously implied, and it changes what can be done about it.

_circlecorner computes raw = (90 - angle)/180 * segs(r, None, fn) and takes ceil(raw). At $fn = 48 a 90° corner has half-angle 45, and (90-45)/180*48 is exactly 12. A ceiling on an exact integer is a knife edge.

Measured: the port prints half=45 raw=12 ceil=12 at %.17g — landing dead on the integer, every call, both families. The reference's arithmetic lands a hair under 45° at some corners, pushing raw fractionally above 12 and the ceiling to 13. On Rectangle: reference envelope 50 vertices, port 48 (13+13+12+12 against 12×4).

Do not try to fix this. Rounding before the ceiling was tried and reverted — it fixes Y exactly and breaks Three-Fin, because Three-Fin has the same asymmetry and the oracle records it. There is nothing to correct on the port's side.

Scale of the effect

30 of 113 accepted cases affected. 83 exact. Only three keys ever breach: SECTION_AREA_MM2 (29), VOLUME_MM3 (30), MASS_G (22) — and volume is area ×100, mass is volume ×density/1000, so each case has one discrepancy reported three times. Worst relative error 2.24e-05.

By profile: Three-Fin 9, Y 7, A Frame 5, Rectangle 5, T 1, Four-Fin 1. None on Equilateral, General Triangle, Square, Diamond or Cross.

Why the test fails when the geometry is fine

At $fn = 48 a chord deviates from its true arc by r(1 − cos 3.75°) = r × 2.1413e-3: 2.68 µm at r=1.25, 4.28 µm at r=2.00. The port and the reference differ from each other by at most ~0.4 µm. Against the 0.01 mm criterion (§3) that is two orders of margin.

The tests fail because VOLUME_MM3 is compared at 1e-4 absolute against a magnitude near 20,000 — demanding 5e-9 relative agreement from discretised geometry. 3x/Y/steel0.79 breaches on VOLUME_MM3 alone while its area passes: the same discrepancy, judged by two wildly different standards by accident of key naming.

The options, and the constraint

Do not edit tolerance in the oracle JSON. It is inside the hashed document; test_integrity_hash covers everything but fixtures_sha256. Editing it breaks that test by design.

The change belongs in test_oracle.py, which is not hashed:

1. Scale-aware bounds for the three discretisation-limited keys, derived from the 0.01 mm criterion and the section perimeter. Most faithful to where the error originates. Recommended.
2. A relative floor — pass if within absolute tolerance or ~1e-4 relative. Simplest; loosens SECTION_AREA_MM2 from 0.001 to ~0.02 mm².
3. Accept 30 known failures. Honest but make test is never green and a real regression hides among them.

CIVICVS decides. The evidence is above; bring him the recommendation and stop.

---

8. Method that has earned its keep

Read the pinned source; do not recall it. BOSL2 was fetched at 92d697c2856de2fed93a33e858068589cefc2898 and read directly. Every function examined had a detail that mattered and that recollection would have got subtly wrong. The same applied to the generators: reading sb-profiles.scad before writing four_x.py confirmed the call signatures rather than inferring them from call sites.

Check the invocation before concluding anything about state. A su that failed identically on two commands read as a repository problem and was an invocation problem. That is the F-027 pattern and it cost a session's opening exchange.

Verify your own arithmetic before shipping it. A claim in PRECISION.md that a 50 mm radius needs facets ≈ 460 was wrong — deviation falls with the square of the segment angle, so the count grows with the square root of radius and the answer is 158. A three-line script caught it.

Mutation-test every suite before landing it. Break the code deliberately and confirm the tests notice. This found real gaps in five of six slices. It also caught a malformed mutation of mine — cutting the cavities twice is idempotent. A surviving mutation is sometimes a bad mutation, not a test gap.

Read-only before write. Every command group where the answer was not certain established the facts first.

One task, one command group, wait for output. Not a menu of next steps. If you find yourself writing "and also", delete it.

---

9. What the project is for

Build the capacity to construct real structures — the reference case is a faceted timber shell — from reclaimed and commodity materials, using whatever fabrication is to hand. The compiler makes the pieces computable, qualifiable, and reproducible by someone who was not present when they were designed.

Codes and permitting are out of scope, deliberately. The project records physical claims, never verdicts. Measure and attest; never adjudicate.

Three artifact classes: members (prismatic, exist — the eleven profiles), nodes (non-prismatic, no representation yet), panels (sheet, none yet).

The output must eventually be sealed manifolds and printable STL. The 2D section is where manifold validity is decided, not downstream in the CAD kernel — is_region_simple is already a build-blocking check, and a self-touching outline extrudes into something untessellatable.

docs/PRECISION.md states what this compiler does not do: no assembly layer, no structural analysis of any kind, prismatic shapes only, no toolpaths. That document is scope-locked to additive and subtractive manufacturing. Requests to widen it should be refused, not accommodated — see its §10.

---

10. Working with this operator

He is precise, keeps excellent records, and will tell you directly when you are wrong — including when you are being unhelpful. Take it at face value; it is accurate and it is not personal.

He decides. Bring evidence and a recommendation, then stop.

He has delegated all coding decisions — structure, algorithms, test design — and does not want to be consulted on them. Still bring him scope, provenance, and anything irreversible on the host. He runs every command, so nothing is autonomous regardless.

When he says he does not understand something, the writing was unclear. Rewrite it shorter; do not explain it again at greater length.

He watches for scope creep and treats it as the primary risk — "the greatest enemy of software is not bugs, it's feature-creep". Documents and code are both held to it. When in doubt, narrow.

---

11. Open questions, none blocking

#	Question	Owner
1	Should wg-pk narrow mynetworks from 10.110.0.0/22 to explicit hosts?	CIVICVS, estate decision (F-025)
2	Backup strategy — USB, IPFS, optical?	CIVICVS
3	Where is the 3+ TB USB disk attached?	CIVICVS
4	Kane Fabric participant mail, send and receive	Cross-project
5	Tolerance model for the three discretisation-limited keys — see §7	CIVICVS, ready to decide

Question 4 is real and unaddressed. Containers do not send mail by standard and nothing can reach vmbr1 from outside, so receiving has no path at all. It needs a design conversation, not a configuration change.

---

12. Commit log

114189c  docs: PRECISION.md -- scope, guarantees and limits, for lay readers
7b3182c  port: the 3x and 4x profile catalogues; build() now exists
af196a2  docs: one canonical handoff, rewritten in place; F-035
009fcce  docs: handoff for the 19 AUG session
86a47c3  rounding: record F-034, arc segment counts tip on the last bit
8d79431  geom: port sb-core and sb-profiles, the N-generic arrangements
b101fe6  geom: port sb-report, validation and report formatting
ebf02d6  geom: port sb-join, the junction and envelope strategies
38ea024  geom: Shapely-backed region layer
545eee7  geom: port BOSL2 round_corners and path cleanup
dfd02a4  geom: port the pure-geometry half of sb-geom
b67cc12  verify.sh: reach the restore on the diff branch
1d3eed0  Handover push
6967eef  Conformance updates.
c7e32d8  Seed repository: rev-8.0.0 reference, frozen oracle, toolchain, test harness