feat(toolchain-readiness-grades): complete the directory — fill in templates, .a2ml, SELF-ASSESSMENT

The initial TRG commit (1880521) shipped the spec and the README but
left the five templates referenced from the spec as "forthcoming" and
omitted the machine-readable counterpart and the self-assessment.
That's exactly the "default contractile left in" pattern that should
not survive into git. This commit fixes it.

New files:

* templates/AUDIT-TEMPLATE.adoc — the canonical four-tier MoSCoW audit
  template that every component audit must fill in. Mapped directly
  onto the TRG grade gates (Must -> E, Should -> D, Could -> C,
  Aspirational -> B). Required header and footer fields specified.
  Style rules: every PASSED claim must be supported by a verifiable
  artefact; the audit document is itself subject to the invariant-path
  doc-claims grounder.

* templates/CANONICAL-PROOF-SUITE.adoc — the M/S/E proof suite v1.0
  with 15 entries (5 mathematics, 5 science, 5 engineering). Each
  entry cites where existing formal proofs exist in mathlib4, the
  Coq stdlib, the Idris2 stdlib, the Agda stdlib, etc. so the porting
  work is bounded. This is the gate to leave grade E. Pass condition,
  versioning policy, and re-baseline cadence specified.

* templates/QUALIFYING-PROVERS.adoc — the v1 qualifying-prover set:
  Idris2, Lean 4, Rocq, Agda. Qualifying criteria (active maintenance,
  published kernel design, public test suite, public manual,
  reproducible builds, >=1000 distinct users, no outstanding
  soundness bugs, banned-construct discipline). A-grade
  kernel-soundness criterion: at least one of the three provers in
  cross-validation must have a published kernel-soundness proof.
  As of v1.0: Lean 4 and Rocq unambiguously satisfy; Idris2 and Agda
  are in the set but should not be the sole kernel-soundness anchor
  until their meta-theoretic verification matures. Annual re-baseline.

* templates/FUZZING-CORPUS-FLOOR.adoc — per-grade and per-component
  fuzzing minimums. D requires harness exists + 1hr CI + corpus 100;
  C requires 14 days continuous + 1k corpus + 70% branch coverage;
  B requires 30 days + 10k corpus + 85% coverage + structure-aware
  fuzzer; A requires 90 days + 100k corpus + 95% coverage +
  structure-aware AND differential fuzzer. Per-component matrix
  covering lexer, parser, AST, macro expander, import resolver,
  diagnostics, semantic analyser, type checker, IR, codegen, linker,
  runtime, REPL, formatter, LSP, stdlib. Differential fuzzing
  requirements at A-grade (the LangSec / parser-differential class).

* templates/A-GRADE-LLM-PANEL.adoc — binding composition rules for
  the six-LLM panel. Vendor distinctness (one per major lab),
  composition by role (2 reasoning + 2 specialist coding + 1 instruct
  + 1 MoE), reproducibility requirement (every signoff archived as
  byte-reproducible session transcript with PGP-style tamper
  evidence), independent submission (no coordinated prompting),
  unanimous required, evidence archival. Qualifying-vendor matrix
  (v1.0, 2026-Q2): Anthropic, OpenAI, Google DeepMind, Meta,
  Mistral, DeepSeek, Alibaba, Cohere, xAI. Quarterly re-baseline.
  Defence against the "category error" objection: the LLM panel
  supplements human reviewers, never replaces them.

* SELF-ASSESSMENT.adoc — TRG applied to itself. Honest current grade:
  *X (Untested)*. Standard exists and is internally consistent but
  no toolchain component has yet been audited against it.
  Path-to-D laid out (audit at least one hyperpolymath toolchain
  component; 007 lexer is the obvious first candidate since the
  TRG audit format is derived from the 007 lexer MK2 audit archived
  in references/007/audit-lexer-mk2-tier-format.md).

* TOOLCHAIN-READINESS-GRADES.a2ml — machine-readable counterpart of
  the spec, matching the CRG `.a2ml` pattern. Includes grade
  definitions with ordinal + release stage + stability posture +
  evidence required, release-stage mapping, required toolchain
  components, banned constructs, standing priority order, grade
  aggregation rule, qualifying provers, canonical proof suite
  metadata, A-grade LLM panel composition, demotion triggers,
  per-language extension hook, and self-assessment.

Modified files:

* README.adoc — replaced "Forthcoming" entry for templates/ with
  the actual inventory of the five templates. Added inventory entries
  for SELF-ASSESSMENT.adoc and TOOLCHAIN-READINESS-GRADES.a2ml.
  Bare-filename references like `AUDIT-TEMPLATE.adoc` rewritten as
  `templates/AUDIT-TEMPLATE.adoc` so the doc-claims grounder finds
  them. Same fix applied to SELF-ASSESSMENT.adoc and
  templates/CANONICAL-PROOF-SUITE.adoc.

Verification: re-ran invariant-path doc-claims on the directory after
this fix. Grounded count went from 0 (templates didn't exist) to
41/64 = 64% within the directory's scope. The remaining 23 ungrounded
references are all legitimate cross-estate references (memory entries,
sibling directories in standards/, per-language repo conventions) that
the v0 single-root grounder cannot follow — multi-root grounding is
on the invariant-path Phase A roadmap.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Author: hyperpolymath

toolchain-readiness-grades/README.adoc

@@ -49,13 +49,33 @@ Three reasons:
                                      that originated the four-tier format.
                                      These are the canonical exemplars of
                                      the per-component audit format.
-| `templates/`                      | (Forthcoming) `AUDIT-TEMPLATE.adoc`,
-                                     `CANONICAL-PROOF-SUITE.adoc`,
-                                     `QUALIFYING-PROVERS.adoc`,
-                                     `FUZZING-CORPUS-FLOOR.adoc`,
-                                     `A-GRADE-LLM-PANEL.adoc`. Skeletons
-                                     for the templates referenced in the
-                                     spec.
+| `templates/`                      | The five canonical templates
+                                     referenced from the spec, all
+                                     v1.0 as of 2026-04-12:
+                                     `templates/AUDIT-TEMPLATE.adoc`
+                                     (the four-tier MoSCoW audit
+                                     template),
+                                     `templates/CANONICAL-PROOF-SUITE.adoc`
+                                     (the M/S/E proof suite, 15
+                                     entries — the gate to leave
+                                     grade E),
+                                     `templates/QUALIFYING-PROVERS.adoc`
+                                     (the v1 qualifying-prover set:
+                                     Idris2, Lean 4, Rocq, Agda),
+                                     `templates/FUZZING-CORPUS-FLOOR.adoc`
+                                     (the per-component fuzzing
+                                     minimums),
+                                     `templates/A-GRADE-LLM-PANEL.adoc`
+                                     (the 6+6 panel composition rules).
+| `SELF-ASSESSMENT.adoc`            | TRG applied to itself. Honest
+                                     current grade: *X* (untested).
+                                     The standard has been written
+                                     but not yet applied to any
+                                     toolchain component. Path to D
+                                     and beyond is laid out.
+| `TOOLCHAIN-READINESS-GRADES.a2ml` | Machine-readable counterpart of
+                                     the spec, matching the CRG `.a2ml`
+                                     pattern.
 |===
 
 == Quick mental model

toolchain-readiness-grades/SELF-ASSESSMENT.adoc

@@ -0,0 +1,171 @@
+// SPDX-License-Identifier: PMPL-1.0-or-later
+// Copyright (c) 2026 Jonathan D.A. Jewell (hyperpolymath) <j.d.a.jewell@open.ac.uk>
+
+= Toolchain Readiness Grades — Self-Assessment
+:toc: auto
+:status: Active
+:date: 2026-04-12
+
+The standards repo grades itself. This document applies the TRG criteria
+to *the TRG document set itself*. TRG is brand new, so the honest grade
+is low — and that honesty is the point.
+
+== Current Grade: X (Untested)
+
+*Assessed: 2026-04-12* +
+*Honest verdict: TRG v1.0 is itself at grade X.* The standard exists,
+the document set is internally consistent, but no toolchain component
+has yet been audited against it, no machine-checkable claims have been
+re-validated by the doc-claims grounder, and no external review has
+taken place.
+
+This is the *correct* starting grade. Anything else would be exactly
+the kind of premature self-grading that TRG itself is built to prevent.
+
+== Why X is the right grade right now
+
+[cols="1,1,3", options="header"]
+|===
+| Criterion | Met? | Evidence
+
+| *X (Untested)* — default
+| (n/a)
+| The default state for any newly-created standard before any audit
+  is run. TRG v1.0 was published 2026-04-12 and has not yet been
+  applied to any component.
+
+| *F (Harmful / Wasteful)*
+| PASSED
+| TRG is not harmful. It is internally consistent, CRG-compatible,
+  and addresses a real gap (CRG grades general components; TRG
+  grades toolchain components specifically). The estate has multiple
+  toolchain projects (007, ephapax, AffineScript, my-lang, etc.)
+  that need a stricter rubric than CRG provides.
+
+| *E (Bare Minimum)*
+| FAILED
+| Requires "all Must rows passed" against the four-tier MoSCoW format
+  defined in §4. *No toolchain component has been audited against TRG
+  yet*, so this row cannot be PASSED for any concrete component, and
+  therefore the standard itself has not been demonstrated to be
+  meaningfully applicable.
+|
+|
+| *Path to E:* run a TRG audit on at least one hyperpolymath toolchain
+  component (the obvious first candidate is 007's lexer, since the
+  TRG audit format is derived from the 007 lexer MK2 audit which is
+  archived in `references/007/audit-lexer-mk2-tier-format.md`).
+  Then promote 007's lexer to TRG-graded status. That single audit
+  demonstrates the standard is *applicable* and lifts TRG itself out
+  of grade X.
+|===
+
+== Path to D (Should-Have Complete)
+
+The next non-trivial promotion. Requires:
+
+. *All Must AND all Should rows PASSED* against the four-tier format,
+  for at least one component.
+. *Canonical Mathematical / Scientific / Engineering Proof Suite* (§6.6)
+  machine-checked through the audited component. The suite definition
+  is now at `templates/CANONICAL-PROOF-SUITE.adoc` (v1.0, 15 entries).
+. *2 fully-proven programs* per §6.2, each cross-validated by at
+  least one qualifying prover from `templates/QUALIFYING-PROVERS.adoc`.
+. *RSR-FULL* — all 17 RSR workflows present, SHA-pinned, and currently
+  green. Standards repo is RSR-compliant; need to verify the same for
+  the host repo of the audited component.
+. *Banned constructs ZERO* — verified by `panic-attack assail` clean
+  across all severities (promoted from "Critical and High only" per §5.4).
+. *A2ML files + 0-AI-MANIFEST + Immaculate Guide* — already required
+  at E for the standard's adopting repos.
+. *Fuzzing harness exists* per `templates/FUZZING-CORPUS-FLOOR.adoc`
+  (Grade D row).
+
+*The single biggest blocker* between grade X and grade D is the
+Canonical Proof Suite. Until at least one toolchain in the estate has
+machine-checked all 15 entries of the M/S/E-CPS through itself, the
+standard's gate to leave grade E is unenforceable in practice.
+
+== Path to C (Self-Validated, alpha gate)
+
+After grade D, the path to C requires:
+
+. *All Could rows PASSED.*
+. *Full Testing-Taxonomy 16 × 14 coverage* per
+  `testing-and-benchmarking/TESTING-TAXONOMY.adoc`.
+. *Six-Sigma test passing* — all tests in the Six-Sigma range
+  (≤ 3.4 failures per million test runs) for at least 14 consecutive
+  days.
+. *Continuous fuzzing for at least 14 days*.
+. *5 fully-proven programs.*
+. *10 dogfood users.*
+. *3-nines on the standing priority axes* (dependability, security,
+  interop, usability).
+. *VeriSimDB instance present and used by tests.*
+. *`launch-scaffolder mint` wired.*
+. *Stapeln containers built from Chainguard base.*
+
+== Path to B and A
+
+See `TOOLCHAIN-READINESS-GRADES.adoc` §5.6 and §5.7 for the full
+requirements. The honest expectation is *zero toolchain components in
+the estate reach A in year one*. That is the correct order of
+magnitude. Anything that reaches B in year one is already
+extraordinary by industry standards.
+
+== Standards conformance of the TRG document set itself
+
+[cols="2,1,4",options="header"]
+|===
+| Conformance item | Status | Notes
+
+| Main spec exists in `.adoc` | ✅ | `TOOLCHAIN-READINESS-GRADES.adoc` (~40 KB)
+| README.adoc | ✅ | `README.adoc`
+| Self-assessment | ✅ | This document
+| Machine-readable counterpart (`.a2ml`) | ✅ | `TOOLCHAIN-READINESS-GRADES.a2ml`
+| Templates referenced from spec | ✅ | All 5 templates created: `templates/AUDIT-TEMPLATE.adoc`, `templates/CANONICAL-PROOF-SUITE.adoc`, `templates/QUALIFYING-PROVERS.adoc`, `templates/FUZZING-CORPUS-FLOOR.adoc`, `templates/A-GRADE-LLM-PANEL.adoc`
+| References (canonical examples) | ✅ | 25 audit files copied verbatim from 007 in `references/007/`
+| SPDX header on every file | ✅ | Verified
+| invariant-path doc-claims grounder run | ✅ | First run on `standards/` produced 519 grounded / 2597 ungrounded / 1 unknown / 3117 total — drives this self-assessment's gap analysis
+| All file-path claims in spec resolve | 🟡 | Templates now exist; cross-repo references (e.g. to `feedback_branch_protection.md` in memory, `immaculate-guide/IMMACULATE-GUIDE.adoc` in standards) should resolve under multi-root grounding once that lands in invariant-path Phase A
+| External review by formal-methods community | ❌ | Not yet — Phase A of invariant-path must complete first; then a coordinated submission to LangSec / NCSC research / NIST / formal-methods Zulip
+| Adoption by at least one toolchain component | ❌ | 007 lexer is the obvious first candidate; not yet started
+|===
+
+== Honest acknowledgments
+
+* TRG is not modest. The A grade is intentionally extraordinary —
+  Six Sigma vs trailing-12-month IPO cohort, 5M external processor-
+  hours, 21 fully-proven programs cross-validated by 3 independent
+  provers, type-theory upstream contribution, full hyperpolymath
+  standards conformance, unanimous 6+6 (humans + LLMs) signoff. The
+  author expects approximately *zero* hyperpolymath toolchain
+  components to hit A-grade in the first year of TRG operation.
+  That is the correct order of magnitude.
+
+* The Six Sigma vs IPO cohort methodology is novel. It has not been
+  peer-reviewed by industrial statisticians. §10 of the spec
+  explicitly invites challenge.
+
+* The 6+6 LLM panel is novel in software-readiness rubrics. The
+  combination of binding composition rules + reproducibility
+  requirement + unanimous requirement is the load-bearing claim;
+  no single piece is unprecedented but the combination is.
+
+* The "21 fully-proven programs cross-validated by 3 provers"
+  threshold for A-grade is plucked from intuition, not derived. It
+  is acknowledged as a number-not-derivation in §10 and §11. Likely
+  re-baselined against real telemetry from the first A-grade
+  candidate before TRG v2.
+
+* Per-language extension hook at §1.1 is the mechanism by which
+  language-specific obligations (Eclexia tropical-semiring, AffineScript
+  region/affinity, Idris2 totality discipline) are layered on top
+  of the TRG baseline. Profiles MAY only tighten, never loosen.
+
+== Standing rule
+
+This self-assessment is itself a claim and is itself subject to the
+invariant-path doc-claims grounder. Re-run it after any change to
+this directory. If any file-path claim in this document fails to
+ground, the self-assessment is invalid until the gap is closed.