docs: session log 2026-07-31 — ugt handover built from the gps3 session record

session_log_20260731.md

@@ -0,0 +1,152 @@
+# Session Log — 2026-07-31: `ugt` handover from R740 context, and solver repair
+
+**Machine:** T420 (`finch`)
+**CWD:** `movefaults_clean` — but **the work product is in `ugt`.** See §6.
+**Scope:** build a context handover for continuing the `ugt` project on the
+R740, using this repo's session logs as the evidence base; then fix the
+defects that survey turned up.
+
+Nothing in this repo's code or Bernese state was touched. This log lives here
+because this is where the source material is, and because §3 is a finding
+about *this* repo.
+
+---
+
+## 1. What was produced
+
+| artifact | where | state |
+|---|---|---|
+| R740 context handover for `ugt` | `ugt/docs/R740_HANDOVER_20260731.md`, ~480 lines | **untracked on purpose** — see §3 |
+| Solver fixes | `ugt` branch `fix/solver-numerics-headless` @ `aa549dc` | committed, **not pushed** |
+| This log | `session_log_20260731.md` | untracked |
+
+---
+
+## 2. Method — and its one deliberate constraint
+
+The R740 was **not** reachable and was **not** contacted. The T420 sits on
+`192.168.1.0/24`; the server is on another subnet with no route. Every
+statement about the machine was therefore derived from documents, and the
+handover tags each claim as LOGGED (observed by a session with a shell on the
+box), UNCONFIRMED (asserted, never checked), or VERIFIED (checked here today).
+That tagging is the point of the document; do not flatten it when editing.
+
+**Scoped to the last 7 days on request**, which turned out to be the entire
+lifetime of Claude Code sessions on that server — 07-28 (install verified),
+07-29 (the on-machine storage/RAID session), 07-30 (smartd, mirror, branch
+reconciliation). `docs/BERNESE_GPS3_HANDOVER.md` (07-22/23, reconstructed from
+phone photos, central diagnosis wrong) was deliberately excluded rather than
+merely caveated.
+
+**Transcripts were mined as well as the markdown.** Searched the session
+`.jsonl` files for `lscpu`, `free`, `nvidia` output. **There is none.** The
+"24 physical cores / 48 threads" figure that two documents rely on has never
+been confirmed by a command, after three days of sessions on the box. It is
+tagged UNCONFIRMED in the handover, with the command to settle it.
+
+---
+
+## 3. Finding about THIS repo — disclosed, acknowledged, unremediated
+
+`alfieprojectsdev/movefaults` is a **public** repository. `RESUME_NEXT.md` is
+tracked in it, and states that `R740_PASS` equals `gps3` — i.e. the sudo
+password of a PHIVOLCS server, published alongside its hostname, its IPv4
+address, the login user, and the Cockpit port.
+
+`scripts/deploy_r740.secrets` is correctly gitignored. **The disclosure is in
+the prose, not the file.**
+
+The machine is LAN-only, so this is not remotely exploitable today. It is
+still a live credential in a public index, and git history means deleting the
+line does not retract it.
+
+**The user was told and has acknowledged and accepted this.** No remediation
+was performed this session. If it is ever actioned, the order is: change the
+password first, then scrub the prose, and treat the old one as burned
+regardless of history rewriting.
+
+Consequence carried into the handover: `ugt` is **also** a public repo, so
+that document deliberately omits the host address, management URL, and any
+credential detail. Accepting the exposure here is not a reason to duplicate it
+there.
+
+---
+
+## 4. `ugt` findings — reading was not enough
+
+The survey found three defects by reading. **Running the code found a fourth
+that fires first, and it invalidates the other three's ordering:**
+
+```
+jax.errors.ConcretizationTypeError: The `length` argument to `scan` expects a
+concrete `int` value ... depends on the values of the arguments total_steps
+and norm_interval
+```
+
+Both scripts carry a plain `@jit`, which traces the step counts, while
+`lax.scan` needs `length` as a concrete Python int. **Neither script had ever
+run.** Fixed with `partial(jit, static_argnums=...)`.
+
+Behind it: `batch_lle(...).block_until_ready()` called on the tuple `vmap`
+returns; float32 default; `plt.show()` on a machine that will be headless.
+
+**The float32 claim was asserted in the survey, then measured.** Over the
+committed 200-point sweep, float32 vs float64:
+
+| | |
+|---|---|
+| max abs difference | **0.0588** |
+| LLE range (float64) | −0.0357 … 0.0738 |
+| difference as fraction of range | **54%** |
+| points changing **sign** | **3 of 200** |
+
+Sign is the chaos / no-chaos verdict. float32 was not adding noise, it was
+changing the answer at three points on exactly the boundary the thesis argues
+about. Verification ran the **full committed sweep**, not a reduced one:
+200 params × 50 000 steps, 3.73 s, dtype float64, figure written. ruff 15 → 12
+findings, all pre-existing.
+
+**Unexpected result worth acting on:** that sweep takes **3.7 s on this
+2-core laptop**. The R740 buys nothing for the current workload. It starts to
+matter only for a 2D (R × ω) sweep or far longer integrations. Decide that
+before moving anything.
+
+---
+
+## 5. Gotchas
+
+- **Cross-repo work from the wrong CWD.** All of §1's output lives in `ugt`
+  while the shell sat in `movefaults_clean`. Every git call used
+  `git -C <abs path>`, and the identity check ran first and printed *both*
+  remotes for contrast. A bare `git commit` here would have been silent and
+  wrong — the exact failure the global rule describes.
+- **Reading a script is not running it.** Four defects, three found by
+  reading, and the one that actually fires first found in seconds by
+  `.venv/bin/python lle_solver.py`. The survey would have shipped an ordered
+  list that was wrong at the top.
+- **`plt.show()` needs the backend chosen before pyplot is imported.** In
+  `ugt`, `plt` is re-exported from `sprott_solver` so import ordering cannot
+  silently defeat it — isort would otherwise reorder it into breakage.
+
+---
+
+## 6. State at end / open decisions
+
+**This repo:** untouched except this log (untracked). No background jobs.
+
+**`ugt`:** on `fix/solver-numerics-headless` @ `aa549dc`, one commit ahead of
+`main`, **unpushed**. Uncommitted `main.tex` (+25/−2, the 2026 literature
+review — never committed, so no clone anywhere has it). Untracked
+`docs/R740_HANDOVER_20260731.md` and `.claude/` (a checkout of
+`solatis/claude-config`; do not `git add`). `figures/` now gitignored. A
+`.venv` was created there with jax/ruff.
+
+Open, all needing the user:
+
+| decision |
+|---|
+| push `ugt` branch / open PR |
+| commit `ugt/main.tex` — **the server has no copy of that work** |
+| `ugt` defects 4–6: `setup_env.sh` markdown-mangled curl URL, `[tool.ruff] select` → `lint.select`, `\section\*` + `[Author, 2024]` in `main.tex` |
+| whether to commit this log and the handover, and with what scrubbing |
+| §3 remediation |