runmarker: the RUN marker carries how far behind the tree is — CI structurally cannot (#205)

.claude/doctrine-watermark

@@ -0,0 +1 @@
+b2a3d4706516189d17533e30c9115d8c65b99d64

CLAUDE.md

@@ -27,7 +27,6 @@ disagree, the README wins and this file is the defect.
 
 | path | contents | its own doc |
 |---|---|---|
-| `./onboard.md` | **the entry point for a coding agent installing this fleet elsewhere** — hand it the one-line bootstrap and it reads this | — |
 | `docs/` | the thesis the rest is derived from | — |
 | `prompts/` | five role prompts (`TEAMLEAD` `ARCHITECT` `DEVOPS` `DX` `DEV`) | `prompts/README.md` |
 | `goals/` | per-role standing goals + the standard they must meet | `goals/README.md` |
@@ -55,37 +54,16 @@ Each is explained in full at the pointer; none of it is restated here.
 
 ## Facts with a measurement date, not standing claims
 
-- ⛔ **~~No CI.~~ FALSE since 2026-08-20 — CI exists and GATES.** `.github/workflows/tools.yml`
-  is present on `main` and on multiple refs; `hermetic suites (gating)` is a **required check**
-  and has blocked a merge. *(re-measured 2026-08-20 at `2effc63`, `git ls-tree -r` over remote refs;
-  #272.)*
-  ★ **The commit that falsified this claim cited it as its justification** — *"this repository
-  had no CI, so 23 instruments and 19 suites had never run"* (`239639a`). ⇒ **A measurement is
-  most likely to be falsified by work it caused**, which is exactly when nobody re-checks it: the
-  author of the fix already knows, and the files asserting it are not in the fix's blast radius.
-  ⚠ The old line carried a date, a SHA, a method, and *"re-measure before relying on either"* —
-  **all four, and it still decayed for ~7 hours.** Dating a claim tells a reader it CAN decay; it
-  does not tell them it HAS. ⇒ `scripts/check-orientation.py` now re-measures this one.
+- **No CI.** Zero workflow or `*.yml`/`*.yaml` files on any local or remote ref.
+  *(measured 2026-08-19 at `c465e8e`, 9 refs, `git ls-tree -r` per ref.)* Nothing here gates on
+  a green check, and opening a PR draws no runner.
 - **One git credential for all panes** — GitHub records the operator for whatever any pane
   posts, so name yourself in the body of anything you file. *(issue #4.)*
 
 ⇒ Re-measure before relying on either. A number without a date is a rumour.
 
 ---
 
-## Onboarding someone else's repository
-
-One line, handed to a coding agent — it is a **pointer, not a copy**, so it stays valid while the
-instructions change under it:
-
-> Clone https://github.com/nForma-AI/nForma-NEXT, then read `./onboard.md` at the root — it contains
-> your full instructions for onboarding me through this tool.
-
-⚠ `./onboard.md` installs **four things, not one**: the recipe, the prompts, `scripts/validate-recipe.py`,
-and the goals. ⛔ Without the validator its validation step establishes nothing and reports no error.
-
----
-
 `scripts/check-orientation.py` asserts that every path named above exists. Run it after moving
 or renaming anything. A map whose pointers rot silently is worse than no map, because it is
 still believed.

docs/FOUNDING-THESIS.md

@@ -32,38 +32,11 @@ reserved  : merge · production/deploy · history rewrite · credential handling
 
 `sendCommand` returned `{"sent": true}` for messages that reached nobody, three distinct ways:
 a pane whose agent had exited, a workspace that silently flipped to another project, and text
-that landed in the input box and was never submitted. ⚠ The third **leaves the text in the box**,
-so the intuitive repair — send it again — double-delivers.
+that landed in the input box and was never submitted.
 
 `lastTransitionAt` advanced in **all three**, so the obvious verifier — compare a pre-send stamp —
 passes on every false case.
 
-**Measured since: five, and the fifth is not on this chain.** A mode is individuated by *which
-arrow fails*, whether the loss is *terminal or transient*, and *what false signal accompanies it*
-— never by which API produced the signal, or the list grows once per transport.
-
-| # | mode | arrow | terminal? | false signal |
-|---|---|---|---|---|
-| 1 | pane's agent had exited | gen→del | terminal | `sent: true` |
-| 2 | workspace silently flipped | gen→del *(wrong target)* | terminal | `sent: true` + `lastTransitionAt` |
-| 3 | text in the box, never submitted | del→consumed | terminal | `sent: true` · also `terminal.inject` → `ok` |
-| 4 | alive, delivered, queued, **not yet read** | del→consumed | ⚠ **transient** | `sent: true` + the pane reads *available* |
-| 5 | a standing instruction edited into a file **the launch path never opens** | gen→del | terminal | ⛔ **none** |
-
-⛔ **Mode 4 does not lose the message**, which is why nothing surfaces it: it is latency wearing
-the appearance of availability. The remedy is a **content-free nudge** — never a re-send, because
-the queued copies are still there.
-
-⛔ **Mode 5 is a PULL failure; 1–4 are PUSH failures.** Nothing was sent, so nothing returned
-true. ★ A wrong token eventually gets checked; **an absent signal gives nobody a reason to look**,
-which is why this one persisted longest — measured, committed prompts had reached **one of eight**
-live sessions. ★★ And its symptom is inverted: an undelivered *rule* does not look like a delivery
-failure, **it looks like disobedience**. Seven roles were not emitting a `STATE:` line nobody had
-ever asked them for. No other mode misattributes its failure to the recipient.
-
-⚠ **The count is of modes OBSERVED.** Mode 5 was invisible for a day because it emits nothing, so
-the absence of a sixth is the absence of a signal — not the absence of a mode.
-
 **Consequence:** an instruction is a tracked object with a lifecycle
 (`queued → delivered → acknowledged → acted-on`), and the terminal state is established by an
 **off-pane effect** (branch pushed, file changed, issue transitioned), not by the substrate's

docs/INSTRUCTION-PRECEDENCE.md

@@ -290,51 +290,8 @@ TEAMLEAD as the competent grantor. ⇒ **AUTHORISED.**
 
 ⚠ **The reservation's stated _reason_ is false in this repository, and the verdict survives it
 anyway — which is the part worth keeping.** Measured while writing this file: **zero** workflow
-files and zero `*.yml`/`*.yaml` across all **19** remote refs, so neither command drew a lease.
-
-⛔ **THAT MEASUREMENT EXPIRED THE SAME DAY. Re-measured at `f7b343f`, 2026-08-20:**
-
-```
-.github/workflows/tools.yml   EXISTS
-first run ever                2026-08-20T07:49:09Z   (pull_request)
-runs since                    100
-on: pull_request              <- gh pr create STARTS A RUNNER
-    push: branches: [main]    <- a push to a non-main branch does not
-```
-
-⛔ **The exposure window, and it is the measurement that matters:**
-
-```
-this claim published on main   2026-08-19 21:15   <- TRUE when written
-CI began running               2026-08-20 07:49   <- FALSE from here
-now                            2026-08-20 14:24
-                               -> 6h35m false while published, across 100 runs
-```
-
-⚠ **And the PR carrying this correction drew a runner itself** — `dev4/ci-premise-inverted`,
-`pull_request`, 14:19:06, success. **The document asserting that PRs draw no runner was
-corrected by a PR that drew one.**
-
-⛔ **My first re-measurement of this said "5 runs". It was `gh run list --limit 5`** — the
-default-truncation defect, in the paragraph about a decayed measurement, by the author who
-filed the truncation finding. ⇒ *A reading is checkable only when the response carries its own
-total*, and `gh run list` does not.
-
-⇒ **The correct split is neither the original clause nor its removal.** The clause said push *and*
-PR were spend; #42 removed both on the no-CI ground; **today the PR half is true again and the
-branch-push half is not.**
-
-★ **This is the rolling-baseline defect landing on a RESERVATION rather than on a number, and
-that is the reason it belongs in this file.** A decayed count yields a stale figure. **A decayed
-reservation yields an unmetered action** — and it is silent in *both* directions: over-restriction
-produces no error signal (`goals/README.md`), and under-restriction produces none either, because
-nothing goes red when an agent spends a runner it was told was free.
-
-⚠ **The example above therefore survives on its own terms and its FOOTNOTE does not.** The verdict
-is still AUTHORISED and still by R2 — TEAMLEAD is the named grantor under any rationale — which is
-exactly the property this example was kept to demonstrate: **the procedure keys on WHO MAY GRANT,
-not on why the thing was reserved.** ⇒ The rationale inverted twice in one day and the resolution
-never moved. That justification was
+files and zero `*.yml`/`*.yaml` across all **19** remote refs. There is no pipeline here, so
+`git push` and `gh pr create` draw no metered lease and are not CI spend. That justification was
 carried over from another estate and did not transfer — [#42](https://github.com/nForma-AI/nForma-NEXT/issues/42)
 established it for ARCHITECT's goal file and re-based the reservation on *reserved because
 TEAMLEAD admits work*; the DEV file still carries the untransferred wording ([#16](https://github.com/nForma-AI/nForma-NEXT/issues/16)'s