TGPSKI · TGPSKI · Aug 5, 2026 · Aug 5, 2026
@@ -7,6 +7,17 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ## [Unreleased]
 
+### Changed
+
+- `assets/tui/` is now vendored from its canonical public upstream,
+  [pane](https://github.com/TGPSKI/pane), and picks up the 2026-08-04
+  `bar_chart` upgrades (aggregation binning, outlier clipping, top-row
+  label collision handling) plus the new `interact.py` module
+  (`prompt_search`, `filter_rows`, `hbar`, `cycle`). Byte-identity is now
+  checkable with `make vendor-check` in a pane checkout; provenance
+  updated in `assets/tui/README.md`, `LINEAGE.md`,
+  `references/phase-03-browser.md`, and `PATTERN.md`.
+
 ## [0.1.0] - 2026-08-04
 
 ### Added

@@ -20,6 +20,7 @@ Everything below is checkable. Dates come from `git log`.
 | 2026-07-30 | **2** — the re-application | `15-trust-repair/eval/watch-matrix.sh`, 128 lines | Proof the pattern ports as rules, not code |
 | 2026-08-03 | — | the `no-out` reattribution | L17 — the screen produces suspicion, not evidence |
 | 2026-08-03/04 | **3** — the live view | `matrix_tui.py` (1,869) + `live.py` (629) | Work in flight; probed liveness; process control |
+| 2026-08-04 | **0b′** — the framework goes public | [pane](https://github.com/TGPSKI/pane) | Generation 0b promoted to its own repository: the canonical upstream every copy now re-vendors from, with the byte-identity check as tooling (`make vendor-check`) instead of an instruction |
 
 Nothing here took a quarter to mature, and the pattern did not begin in an
 eval. It began in operations.
@@ -31,16 +32,20 @@ over live server data: `sh-web-analytics` (nginx access logs across several
 domains) and `sh-github-analytics` (repository traffic). Both predate the eval
 suite by more than two weeks.
 
-Both currently live in a private operations repository. `sh-github-analytics`
-is slated for public release as **`catnip`**; this document will carry the link
-once it lands rather than pointing at a URL that does not resolve yet.
-`sh-web-analytics` reads production access logs and is expected to stay
-private.
-
-So for now, unlike every other reference in this document, generation 0 cannot
-be followed. The claims made about it below are deliberately the ones that
-survive that: line counts, dates, a docstring, and a byte-for-byte diff against
-code you *can* read in [leather](https://github.com/TGPSKI/leather) and
+Both apps currently live in a private operations repository.
+`sh-github-analytics` is slated for public release as **`catnip`**; this
+document will carry the link once it lands rather than pointing at a URL that
+does not resolve yet. `sh-web-analytics` reads production access logs and is
+expected to stay private.
+
+The *framework layer* they share, however, is public as of 2026-08-04:
+[**pane**](https://github.com/TGPSKI/pane) is the canonical upstream of
+`shared/tui/`, carrying the provenance hashes of the extraction in its own
+LINEAGE.md. The apps — the readers, views, and callouts — are what remain
+private; the claims made about them below are deliberately the ones that
+survive that: line counts, dates, a docstring, and a byte-for-byte diff
+against code you *can* read in [pane](https://github.com/TGPSKI/pane),
+[leather](https://github.com/TGPSKI/leather) and
 [adherence-suite](https://github.com/TGPSKI/adherence-suite).
 
 They matter for three reasons.
@@ -52,7 +57,7 @@ those two apps on 2026-07-13 — its own docstring says so:
 > bounds-checked put, base color pairs, run loop with min-size guard, footer
 > renderer, scroll indicator, CSV loading, and the curses.wrapper bootstrap.
 
-All four files are **byte-identical** to the `tui/` package vendored into
+All four files were **byte-identical** to the `tui/` package vendored into
 leather's eval scripts and again into adherence-suite:
 
 ```
@@ -62,6 +67,12 @@ framework.py   108 lines   IDENTICAL
 windows.py      58 lines   IDENTICAL
 ```
 
+(Since 2026-08-04 the upstream is [pane](https://github.com/TGPSKI/pane),
+`charts.py` has grown aggregation binning and outlier clipping, a fifth
+module `interact.py` joined at its third duplication, and the identity
+claim is a command — `make vendor-check` in pane — rather than a table
+that can go stale like this one did.)
+
 So **L8's `N–M of T` indicator is not an eval idea.** It is
 `TuiApp.scroll_indicator`, written for a log dashboard sixteen days earlier,
 and every later watcher inherited it without rewriting a line.

@@ -224,10 +224,10 @@ What transfers: rank by the question, name your callouts after what they
 already ruined, distinct glyphs for distinct absences, probe liveness, one data
 path.
 
-**The exception is the drawing layer.** `shared/tui/` — bounds-checked put, a
-min-size guard, a footer renderer, a scroll indicator — has been vendored
-byte-identically into three codebases across twenty-five days, because it
-encodes nothing about any particular job. That is the test for what belongs in
+**The exception is the drawing layer.** [pane](https://github.com/TGPSKI/pane)
+— bounds-checked put, a min-size guard, a footer renderer, a scroll indicator
+— has been vendored byte-identically into four codebases across twenty-five
+days, because it encodes nothing about any particular job. That is the test for what belongs in
 a framework: if it knows what the numbers *mean*, it is not framework, and
 copying it is how a watcher inherits the wrong question.
 

@@ -1,7 +1,9 @@
-# `assets/tui/` — the drawing layer, to be copied verbatim
+# `assets/tui/` — the drawing layer, vendored from pane
 
-Three stdlib-only modules, 326 lines. **Copy them into the target repository in
-Phase 3. Do not rewrite them, and do not improve them.**
+Five stdlib-only modules, ~570 lines, vendored byte-identically from
+[**pane**](https://github.com/TGPSKI/pane) — the canonical upstream of this
+package. **Copy them into the target repository in Phase 3. Do not rewrite
+them, and do not improve them.**
 
 This is the one part of a run watcher that ports as *code* rather than as
 rules, and the reason is the test worth remembering:
@@ -10,33 +12,45 @@ rules, and the reason is the test worth remembering:
 
 Nothing in here knows what a trial, an arm, a request or a repository is. It
 can put a string inside the terminal bounds, refuse to draw in a window too
-small to be honest, render a bar, and say `1-22/73`. That is all — which is
-exactly why it has survived three codebases unchanged while every layer above
-it was rewritten from scratch each time.
+small to be honest, render a bar, prompt for a search, and say `1-22/73`.
+That is all — which is exactly why it has survived four codebases unchanged
+while every layer above it was rewritten from scratch each time.
 
 ## Provenance
 
-Extracted on 2026-07-13 from two live server-analytics dashboards, and vendored
-byte-identically since. The files here match, to the byte, the copies in:
+Extracted on 2026-07-13 from two live server-analytics dashboards, vendored
+byte-identically since, and promoted to its own repository —
+[pane](https://github.com/TGPSKI/pane) — on 2026-08-04. The files here match,
+to the byte, `src/pane/` upstream and the copies in:
 
 - [leather](https://github.com/TGPSKI/leather) — `examples/14-sig-triage/eval/scripts/tui/`
 - [adherence-suite](https://github.com/TGPSKI/adherence-suite) — `src/adherence/tui/`
 
-Verify it yourself rather than believing it:
+Verify it yourself rather than believing it — pane ships the checker:
 
 ```bash
-diff assets/tui/framework.py /path/to/leather/examples/14-sig-triage/eval/scripts/tui/framework.py
+cd pane && make vendor-check    # diffs every portfolio copy against src/pane
 ```
 
-See [LINEAGE.md](../../LINEAGE.md) for why that matters.
+To refresh this copy after an upstream change:
+
+```bash
+cd pane && tools/vendor.sh ../run-watcher/assets/tui
+```
+
+See [LINEAGE.md](../../LINEAGE.md) here and
+[pane's LINEAGE.md](https://github.com/TGPSKI/pane/blob/main/LINEAGE.md) for
+why the byte-identity claim matters.
 
 ## What each module gives you
 
 | File | Provides | Laws it carries |
 |---|---|---|
 | `framework.py` | `TuiApp` (bounds-checked `_put`, colour pairs, run loop, footer, **`scroll_indicator`**), `curses_main` | **L8** — the `N–M of T` indicator is `scroll_indicator`, not something you write per project |
-| `charts.py` | `bar_chart` and friends | — |
+| `charts.py` | `bar_chart`: single or stacked series, peak markers, half-blocks, aggregation binning (long series bin rather than fall off the right edge), outlier clipping (`clip_ratio` caps the y-axis at a robust bound and labels what ran past it) | **L5**'s spirit at the axis: the y-max must describe a bar the user can actually see |
 | `fmt.py` | `compact_num`, `human_bytes`, `duration`, `pct`, `sparkline`, safe `to_int`/`to_float` | **L5** — the coercers return a default rather than raising on a torn value |
+| `interact.py` | `prompt_search` (bottom-row `/`), `filter_rows`, `hbar`, `cycle` | Extracted at the third duplication; takes the app as an argument so `framework.py` stays byte-identical |
+| `windows.py` | `trailing_hours`, `trailing_days`, `trend`, `stack_cells` | see below |
 
 Two things in `framework.py` are load-bearing and easy to delete by accident:
 
@@ -48,7 +62,7 @@ Two things in `framework.py` are load-bearing and easy to delete by accident:
 
 ## `windows.py`, and why it ships anyway
 
-All four modules are here, including `windows.py`, which looks at first like
+All five modules are here, including `windows.py`, which looks at first like
 pure analytics vestige — `trailing_hours`, `trailing_days` and `trend` are
 time-series helpers whose docstring names the "hourly CSV wire format", and no
 watcher in three generations has called any of them.
@@ -64,10 +78,8 @@ if segments:
 time-series module. Dropping `windows.py` leaves `bar_chart` working for every
 call the existing watchers make, and raising `ImportError` the first time
 someone passes stacked segments — a latent failure planted in the one component
-that is supposed to be the safe part.
-
-Keeping the package whole also keeps it byte-identical to the public copies,
-which is what makes the `diff` above worth running.
+that is supposed to be the safe part. This is why pane's `tools/vendor.sh`
+copies the whole package or nothing.
 
 **Genuinely unused, and safe to ignore:** `read_csv` in `framework.py`, and the
 three time-series functions above. They survive for byte-identity, not because
@@ -77,6 +89,13 @@ The lesson generalizes past this file: **"unused" and "unreachable" are not the
 same claim.** A lazy import inside a branch is invisible to every check that
 reads the top of a file.
 
+## Testing what you build on it
+
+pane also ships `pty_smoke.py` (not vendored here — it's test infra, and
+POSIX-only): drive any TUI in a real pseudo-terminal, feed it keys, assert a
+clean exit with no traceback. Point it at your watcher from your own test
+suite rather than writing a second harness.
+
 ## What you still have to write
 
 Everything that knows anything:
@@ -87,5 +106,5 @@ Everything that knows anything:
 - every law from the [design laws](../../references/design-laws.md) except L8's
   indicator, which you get for free
 
-The framework is roughly 300 lines. A useful watcher is 300–2,500 more. Copying
+The framework is roughly 570 lines. A useful watcher is 300–2,500 more. Copying
 this is the last shortcut you get.
@@ -0,0 +1,9 @@
+"""Vendored curses-TUI primitives (stdlib only).
+
+Source: https://github.com/TGPSKI/pane @ a51e682
+(src/pane: framework.py, charts.py, fmt.py, windows.py, interact.py),
+copied verbatim so this repository runs from a fresh clone with no
+external path dependency. Upstream owns the API; keep edits there and
+re-vendor with pane's tools/vendor.sh rather than diverging here.
+Verify with pane's tools/check-vendor.sh.
+"""