Merge: link skills to their characteristic, and make the canonical path the default

A skill names its characteristic only inside <CHARACTERISTIC_CHOICE>. The HDT
provider read that block for its costs and dropped the name, so a document that
omits CHARACTERISTIC produced a general-level roll — a flat 11 — instead of
9 + characteristic/5 (6E1 p57). That path is kirby-api's relational
re-emission and the in-app editor; every HD-saved .hdc states the
characteristic, so the corpus never saw it.

Also bundles the three authored characters and retires the env-var gate that
kept the canonical loader out of the default run. Fourteen tests were skipping
on one unset variable while sixteen dump-based tests stood in for them — and
the dump is lossy on exactly the two things this change touches, which cost us
two false alarms before anyone loaded the real file.

Oracle parity held at 656/656 with the residual ledger untouched. Suite goes
1732 passed / 14 skipped to 1753 passed / 0 skipped.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01PrVDXgpDfQpzEGErZTLjEs

Author: pdbethke

.env.test.example

@@ -43,10 +43,10 @@ KIRBY_COST_HD6CLI=~/path/to/hd6cli.sh
 # artifact to name and no default worth writing.
 KIRBY_COST_ROUNDTRIP_HDC=~/path/to/a-complex-character.hdc
 
-# A directory holding the three authored characters cleared for use as
-# corpus, by name: Ravel.hdc, Bokor.hdc, PowerLad.hdc. Symlinks are fine —
-# they need not live together on disk. The .hdc files are NOT redistributed;
-# only the oracle's JSON dumps of them ship, and those carry costs and
-# display strings with no campaign material in them. Unset: the three
-# authored-character tests skip.
+# OPTIONAL, and an override rather than a requirement since 2026-09-01: all
+# three authored characters (Ravel.hdc, Bokor.hdc, PowerLad.hdc) now ship in
+# tests/fixtures/authored/, so nothing skips when this is unset. Set it to a
+# directory holding those three names to run against your own working copies
+# instead — symlinks are fine, they need not live together on disk. A name
+# this variable does not resolve falls back to the bundled copy.
 #KIRBY_COST_AUTHORED=/path/to/authored

.gitignore

@@ -72,4 +72,27 @@ legal/
 # roundtrip_hd6_costs.json is one character's costed build. Same rule.
 tests/fixtures/oracle/
 tests/fixtures/*.hdc
+tests/fixtures/**/*.hdc
 tests/fixtures/roundtrip_hd6_costs.json
+
+# DELIBERATELY TRACKED, and the only .hdc files in the repository:
+# tests/fixtures/authored/{Ravel,Bokor,PowerLad}.hdc are PeterB's own
+# characters, built on stock templates and cleared by him for publication
+# (2026-09-01). They are here so the canonical load path -- HDCLoader on a
+# real HD-saved file -- runs by default instead of only where an environment
+# variable happened to be set.
+#
+# The `tests/fixtures/**/*.hdc` line above is what makes these three
+# negations load-bearing, and it was added for that reason. `*.hdc` alone
+# does not match a subdirectory, so before it every .hdc under
+# tests/fixtures/authored/ was tracked by default -- including any licensed
+# third-party character dropped there, which `git add -A` would have staged
+# without a word. Deny everything under tests/fixtures/, re-admit exactly
+# three by name.
+#
+# THREE FILES, NO MORE. Nothing else from that document store may be added:
+# the rest is licensed third-party content and personal campaign material,
+# including the variants that sit beside these (background sheets, kits).
+!tests/fixtures/authored/Ravel.hdc
+!tests/fixtures/authored/Bokor.hdc
+!tests/fixtures/authored/PowerLad.hdc

.superpowers/canonical-fixture-report.md

@@ -0,0 +1,277 @@
+# Canonical fixture correction — `fix/skill-characteristic-linkage`
+
+*2026-09-01*
+
+## The question
+
+"Does not canonical imply default?" It did not. `HDCLoader` is this project's
+canon — 100% Java-oracle parity, and the only path that reads what HERO
+Designer actually wrote. But the skill-roll tests on this branch asserted
+against `tests/fixtures/authored/Ravel.json`, an **oracle dump**, and the one
+test that loaded through the canonical path was gated on `KIRBY_COST_AUTHORED`,
+which nothing set. Sixteen dump-based tests ran; the faithful one did not.
+
+The dump is lossy in two ways that matter:
+
+1. It omits `CHARACTERISTIC` on every skill.
+2. It does not record which ITEM a multi-ITEM skill (`PROFESSIONAL_SKILL`,
+   `SCIENCE_SKILL`) was bought as, so the engine can only take the template's
+   first.
+
+That produced two false alarms in one evening — five "proficiency regressions"
+and two "multi-ITEM build-doc gaps" — both of which evaporate on a real load.
+
+## What changed
+
+**1. The characters are bundled.** `tests/fixtures/authored/` now holds three
+`.hdc` files, verified by `CHARACTER_NAME` before copying:
+
+| file | bytes | `CHARACTER_NAME` |
+|---|---|---|
+| `Ravel.hdc` | 172,400 | `Ravel, The Unmade Man` |
+| `Bokor.hdc` | 140,178 | `Bokor` |
+| `PowerLad.hdc` | 85,944 | `Power Lad` |
+
+md5 verified identical to the originals. **Nothing else was copied** from
+`~/Documents/Champions` — not `Ravel_background.hdc`, not `Ravel (CSI Kit).hdc`,
+not any licensed third-party character. `.gitignore` gained explicit
+`!` negations for exactly these three plus a comment saying three files, no
+more, so tightening `tests/fixtures/*.hdc` later cannot silently drop them and
+a fourth cannot be added by accident.
+
+**2. `authored_hdc()` finds them with no environment variable.**
+`tests/corpus.py` gained `BUNDLED_AUTHORED`; `authored_hdc(name)` checks
+`KIRBY_COST_AUTHORED` first and falls back to the bundle. The override is
+per-name, so a maintainer pointing at his own store still gets his working
+copies, and a name that store lacks still resolves.
+
+**3. A missing canonical fixture FAILS.** New `require_authored_hdc(name)`
+raises `FileNotFoundError` naming the bundled path. The canonical tests carry
+**no skip guard at all** — in `test_skill_characteristic_roll.py`,
+`test_campaign_cost_fields.py`, and `test_authored_characters.py`'s `_cases()`.
+These files are tracked, so absence is a deletion, not an unconfigured machine.
+
+**4. The skip guard was ARMED, not bypassed.** `conftest.py`'s design is
+untouched: it still reads `INPUTS` twice, to report what a run was configured
+with and to decide whether a skip is acceptable or a defect.
+`KIRBY_COST_AUTHORED` stays in `INPUTS` — it is still honoured as an override.
+What changed is `missing_inputs()`: via `_BUNDLE_SATISFIES`, an input the
+repository itself carries is no longer "missing". That tightens the guard
+rather than loosening it. The run header on this machine now reads:
+
+```
+kirby-cost: all 9 inputs present — any skip will fail the run
+```
+
+Before, it reported `KIRBY_COST_AUTHORED` missing and the guard excused all 14
+skips. Three tests in `test_test_inputs.py` pin the new contract (bundle
+satisfies; variable still overrides; missing bundle raises).
+
+**5. The tests moved to the canonical path.** `test_skill_characteristic_roll.py`
+is now explicitly two-lane, and says which lane each test is in:
+
+- **Canonical** — the `ravel` fixture is `HDCLoader().load_file(...)`,
+  function-scoped and reloaded per test on purpose, because `load_file`
+  *installs the active hero* and `roll_value`'s characteristic branch reads it.
+  Caching it would be correct only until another test loaded another character.
+- **Degraded, deliberately** — the `ravel_dump` fixture builds from the JSON
+  dump, and is the only shape in which the template fallback fires at all
+  (across the 655-character corpus it fires on zero of 4,434 skill-like
+  objects). `test_the_dump_states_no_characteristic_at_all` pins that premise,
+  so the fallback tests cannot go on passing while testing nothing. The
+  familiarity/Everyman short-circuit tests keep building from a mutated doc,
+  labelled as such.
+
+**6. The test that would have caught tonight.**
+`test_ravel_reproduces_every_roll_hero_designer_printed` loads Ravel
+canonically and asserts sixteen skill rolls against HD's own rendered output.
+`test_the_transcribed_rolls_are_still_the_ones_hero_designer_printed`
+re-derives that table from the dump's `column2_output` so the transcription
+cannot rot into a test of a typo. Seven of the sixteen — the five Proficiencies
+and both multi-ITEM skills — are precisely the cases a dump-based test cannot
+state.
+
+## Ravel against HD's printed rolls, via the canonical loader
+
+Every skill for which HD renders a roll. **16 of 16 agree, 0 mismatched.**
+
+| skill | characteristic | engine | HD printed |
+|---|---|---|---|
+| PROFESSIONAL_SKILL | INT | 14 | 14 |
+| SCIENCE_SKILL | INT | 14 | 14 |
+| DEDUCTION | INT | 14 | 14 |
+| CRIMINOLOGY | INT | 14 | 14 |
+| FORENSIC_MEDICINE | INT | 14 | 14 |
+| PARAMEDICS | INT | 14 | 14 |
+| COMPUTER_PROGRAMMING | INT | 14 | 14 |
+| NAVIGATION | INT | 14 | 14 |
+| ACTING | PRE | 13 | 13 |
+| CHARM | PRE | 13 | 13 |
+| KNOWLEDGE_SKILL | GENERAL | 13 | 13 |
+| HIGH_SOCIETY | PRE | 10 | 10 |
+| BUREAUCRATICS | PRE | 10 | 10 |
+| SECURITY_SYSTEMS | INT | 10 | 10 |
+| STREETWISE | PRE | 10 | 10 |
+| INTERROGATION | PRE | 10 | 10 |
+
+`LANGUAGES` and `SKILL_LEVELS` are excluded because HD renders no roll for
+either, and the provenance test asserts that exclusion set exactly.
+
+The five Proficiencies sit at 10 despite PRE/INT linkage: `roll_value`
+short-circuits on the proficiency flag before it consults the characteristic
+(6E1 p57 sets the linked base; a Proficiency is bought as a flat roll).
+
+## Gates
+
+**Oracle (release gate):** `656 passed`, exit 0.
+`tests/fixtures/oracle_known_residuals.json` untouched — `git status` reports
+no modification to it.
+
+**Full suite**, `KIRBY_COST_HDT` pointed at the complete 17-template directory:
+
+| | passed | skipped |
+|---|---|---|
+| before (stashed working tree) | 1732 | **14** |
+| after | **1753** | **0** |
+
+The 14 skips were all `KIRBY_COST_AUTHORED unset`: 12 in
+`test_authored_characters.py` (4 parametrized tests x 3 characters), 1 in
+`test_campaign_cost_fields.py`, 1 in `test_skill_characteristic_roll.py`.
+All 14 now run. **Zero skips remain.**
+
++21 net passing: 12 previously-skipped authored tests + 2 previously-skipped
+canonical tests, plus the new canonical and provenance tests in
+`test_skill_characteristic_roll.py` and three new guard tests in
+`test_test_inputs.py`, less the dump-based tests that were folded into the
+canonical ones.
+
+## The caution that did not fire
+
+`test_authored_characters.py` compares each `.hdc` against its JSON fixture on
+every cost and every display string, for all three characters. Those twelve
+tests had never run in this environment. **They pass on first run** — no
+disagreement between the real files and the dumps on anything those tests
+compare. That is a genuinely useful negative result: the dumps are faithful on
+*costs and strings*, and lossy specifically on the **skill characteristic
+linkage and multi-ITEM selection**, which is the narrow seam this branch is
+about. Nothing was adjusted to make anything pass.
+
+## Not touched
+
+`Hero.characteristic_value()` — unchanged.
+`tests/fixtures/oracle_known_residuals.json` — unchanged.
+`conftest.py` — unchanged; the guard mechanism it implements is now armed by
+default rather than modified.
+
+---
+
+# Fix round — the licensing rail was inert
+
+*2026-09-01, after review*
+
+## What was wrong
+
+The three `!tests/fixtures/authored/*.hdc` negations I added were **inert**,
+and worse than inert: they were a claim of protection where none existed.
+
+The pre-existing pattern is `tests/fixtures/*.hdc`. A single `*` does not
+cross a directory separator, so that line never reached
+`tests/fixtures/authored/` at all. Nothing there was ever ignored, so
+negating it re-admitted nothing. The three bundled files stayed tracked for
+the same reason they would have without any of my lines: they were never
+matched.
+
+The consequence is the part that mattered. A licensed third-party `.hdc`
+dropped into `tests/fixtures/authored/` would have been picked up by
+`git add -A` with no warning — the exact failure the comment above it claimed
+to prevent. My report said a fourth file "cannot be added by accident" and the
+SKILL.md said the three were "explicitly un-ignored". Both described a
+mechanism that was not running. A false assurance on the rail that keeps
+licensed content out of a public repo is worse than no claim at all, because
+it stops anyone from checking.
+
+## The fix
+
+One line, above the negations, which is what makes them load-bearing:
+
+```
+tests/fixtures/oracle/
+tests/fixtures/*.hdc
+tests/fixtures/**/*.hdc      <-- added
+tests/fixtures/roundtrip_hd6_costs.json
+...
+!tests/fixtures/authored/Ravel.hdc
+!tests/fixtures/authored/Bokor.hdc
+!tests/fixtures/authored/PowerLad.hdc
+```
+
+Deny every `.hdc` anywhere under `tests/fixtures/`, then re-admit exactly
+three by name. The surrounding comment now says *why* the `**` line exists and
+tells the next reader to verify with `git check-ignore -v` rather than trust
+the prose — since trusting the prose is what went wrong here.
+
+## Proof, run after the change
+
+```
+$ git check-ignore -v "tests/fixtures/authored/Ravel_background.hdc"
+.gitignore:75:tests/fixtures/**/*.hdc	tests/fixtures/authored/Ravel_background.hdc
+exit=0
+
+$ git check-ignore -v "tests/fixtures/authored/Ravel (CSI Kit).hdc"
+.gitignore:75:tests/fixtures/**/*.hdc	tests/fixtures/authored/Ravel (CSI Kit).hdc
+exit=0
+
+$ git ls-files tests/fixtures/ | grep hdc
+tests/fixtures/authored/Bokor.hdc
+tests/fixtures/authored/PowerLad.hdc
+tests/fixtures/authored/Ravel.hdc
+
+$ git status --porcelain
+ M .gitignore
+ M .superpowers/canonical-fixture-report.md
+ M tests/test_skill_characteristic_roll.py
+```
+
+Both licensed names are now ignored, and `.gitignore:75` names the line doing
+it. Exactly the three bundled files remain tracked. The porcelain output shows
+only this fix round's own edits and no stray `.hdc`.
+
+**Live drill, not just a dry check.** I copied the real
+`Ravel_background.hdc` into `tests/fixtures/authored/`, ran `git add -A .`,
+and confirmed the staged set:
+
+```
+$ git diff --cached --name-only | grep hdc
+NONE STAGED
+```
+
+The file was physically present in the directory and `git add -A` did not
+stage it. Removed afterwards; the directory holds the three bundled `.hdc`
+and their four `.json` fixtures again.
+
+## Minors
+
+- "fifteen skill rolls" corrected to **sixteen** in both places in this
+  report. `HD_PRINTED_ROLLS` has sixteen entries and the table above lists
+  sixteen rows; the prose was simply wrong.
+- `test_the_transcribed_rolls_are_still_the_ones_hero_designer_printed` no
+  longer takes the `ravel` fixture. It reads HD's rendered output and the
+  transcription and never loads a character, so the parameter was buying an
+  `HDCLoader` run per invocation for nothing. Its docstring now says it takes
+  no hero, so the parameter does not come back.
+- SKILL.md's "explicitly un-ignored" wording replaced. It now states what is
+  actually true — everything under `tests/fixtures/` is denied, exactly three
+  are negated back in — names the `**` line as the one that reaches
+  subdirectories, and records that negations against `*.hdc` alone were inert,
+  so the next person does not repeat this.
+
+## Gates re-run after the fix
+
+| | result |
+|---|---|
+| Oracle | **656 passed**, exit 0; `oracle_known_residuals.json` unmodified |
+| Full suite | **1753 passed, 0 skipped**, exit 0 |
+
+Unchanged from before the fix round, as expected: nothing here touches engine
+behaviour, and dropping an unused fixture parameter removes a redundant load
+rather than a test.