# Pinned workflow declaration units

This is a non-Markdown workflow declaration inventory, not a CI/CD presentation
or evidence that the workflows have run. It uses the presentation source pin
`1ceae52c7896a59ac74a4429d410f3218f0b91c6` and reconciles its complete selection
with both Git and the [whole-repository discovery](../repository/README.md).

All **106 direct `.github/workflows/*.yml` / `*.yaml` files** and **two tracked
`action.yml` / `action.yaml` definitions** are included: **1,065,561 source
bytes, 308 jobs, 1,767 steps and 3,718 meaningful review units**. All parse
under the explicit supported subset; none is omitted for being old, repetitive,
generated or outside the existing presentation citations. This does not select
nested workflow-looking files, other CI systems or scripts by filename
resemblance.

| Unit role              | Count | Review obligation                                                                                     |
| ---------------------- | ----: | ----------------------------------------------------------------------------------------------------- |
| Workflow setting       |   452 | Trigger/filter, permission, concurrency, environment or other complete top-level declaration          |
| Job setting            | 1,383 | Runner, dependency, condition, strategy/matrix, environment, output or other complete job declaration |
| Step                   | 1,767 | Entire ordered step, including its command/action, arguments, conditions and local error policy       |
| Action setting         |     6 | Complete local action metadata, input and output declarations                                         |
| Action runtime setting |     2 | Declared runtime configuration outside composite steps                                                |
| Source context         |   108 | Remaining comments, container identities and syntax; not automatically exempted as structural         |

Nested mappings, matrices and multiline scripts remain intact inside their
declaration or step. The collector does **not** parse individual shell
statements, resolve expressions, execute tools, install packages, fetch remote
actions or assert effective workflow validity. Those deeper implementation and
operational reviews remain required. The declarations contain 326 literal
job-`needs` references; none names an absent job in its own file or needs
expression resolution in this projection. This is not cycle detection,
scheduling evidence or proof that a success job enforces every other job's
result.

## Identity, exact bytes and privacy

Each source records its Git blob ID and whole-file SHA-256. Each unit records a
stable source-path/YAML-location ID, role, exact half-open UTF-8 byte ranges,
inclusive source lines and hashes. A unit hash is SHA-256 of the canonical JSON
array of its ordered range SHA-256 values. Range coordinates are not in that
hash. Across a parsed file the ranges partition **every source byte exactly
once**, including comments and trailing whitespace; the source-context unit owns
the complement of the explicit declarations and steps.

Unit IDs exclude line numbers and content. Unrelated inserted lines preserve an
existing declaration identity. Step identities are positional (`steps/0`, etc.);
insertion or reordering requires review, not automatic transfer of acceptance.
Repeated steps and identical files retain every context. A reviewed assignment
also binds the **whole source file**, so a changed trigger, inherited
permission, environment or matrix invalidates dependent explanations even if a
step's own bytes remain unchanged.

Only structural keys, job identifiers/needs and literal local-action paths are
projected. General scalar values, command bodies, names/descriptions, input
values, credential expressions and remote action reference values are not copied
into the inventory. Remote reference hashes support identity comparison; read
the pinned source for the actual contract. Local action definition candidates
mean tracked presence at the pin, not the runner's checkout or runtime
resolution. They do not prove that a local action was invoked safely.

Parsing uses PyYAML's non-constructing `BaseLoader` node representation,
retaining the literal `on` key. Duplicate/complex keys, aliases/anchors, merges,
explicit unsupported tags, multiple documents, malformed encodings and
size/depth/event limit violations become explicit source issues. Their file
identities remain accounted for, but no partial unit set can masquerade as a
complete parse. Parser diagnostics use fixed codes rather than echoing possibly
private source.

## Teaching assignments and remaining work

All **3,718 units are unassigned**. Ownership, responsibilities, implementation
state, external/local integration, substantive explanations and runtime evidence
still need source review. No generic slide, source dump, directory listing or
unit count establishes that teaching has happened.

Future `workflow-assignments/*.json` files use an `assignments` array. Each
entry needs `unitId`, the unit's `sourceSha256`, the source's
`sourceFileSha256`, `mode: "explained"`, `reviewed: true`, a substantive
`rationale`, and `evidence`. Each evidence item names a real `guideId`,
`slideId`, the assembled `slideSha256`, and `explains`. The slide must belong to
that guide and cite the YAML source through `sources`, without an unrelated
Markdown heading filter. Hashes validate an editorial assertion; they cannot
decide whether the explanation is genuinely faithful or complete. No assignments
are fabricated for this collection.

`check-workflow-coverage.py` rejects unknown/repeated units, missing review
evidence, source/file-context/slide staleness, invalid membership/citations and
unsupported source issues. `--require-complete` additionally rejects unassigned
or merely assigned units. Its success would cover **this declaration family
only**, never all repository code/configuration/assets or actual CI execution.
The documentation-unit gate remains independently required and incomplete.

Next: review triggers and job results against the actual scripts/actions; build
source-backed orientation, gate, artifact, deployment and recovery explanations;
trace environment/permission/credential boundaries; add worked passing, skipped,
cancelled and failing paths; then bind reviewed guide/slide evidence. Expand
semantic accounting to the remaining code, schema, workflow and asset families.

## Reproduction

Use `tools/presentations/requirements.txt` and the repository's resource-aware
workflow. No browser, model or remote CI is launched by these commands.

```sh
python tools/presentations/inventory-workflows.py
python tools/presentations/inventory-workflows.py --check
python tools/presentations/check-workflow-coverage.py
python tools/presentations/check-workflow-coverage.py --check
python tools/presentations/check-workflow-coverage.py --check --require-complete
python -m unittest discover -s tools/presentations/tests -p 'test_workflow_inventory.py'
node docs/presentations/presentation-center/verification/workflow-inventory-2026-09-07/verify-evidence.mjs
```

The complete-teaching command is **expected to exit 1** today. Freshness is a
separate gate: `--check` recollects exact source units and checks all artifact
bytes. Unknown or obsolete JSON files fail rather than being deleted. The
[verification receipt](../../verification/workflow-inventory-2026-09-07/README.md)
records the actual checks and the independent all-byte audit.
