name	grind-err
description	Systematically reduce stzlib

Grind ERR

A repeatable loop for reducing #ERR-tagged test failures in libraries/stzlib/base/test/<topic>/. The point of the skill is discipline -- the same checks every time so a stray duplicate def or arity collision doesn't masquerade as a regression.

When to invoke

"grind string", "grind char", "grind "
"continue the grind", "keep going on path 1"
"reduce ERR count", "burn down failures"
After landing a fix, when the user says "continue"

The loop (one iteration)

1. Baseline

TOPIC=string  # or char, dataset, etc.
DIR=libraries/stzlib/base/test/$TOPIC

# live count of files carrying #ERR headers
grep -l '^#ERR' $DIR/*.ring | wc -l

# error-type distribution (the leverage map)
grep -h '^#ERR' $DIR/*.ring | sort | uniq -c | sort -rn | head -10

2. Pick the highest-leverage cluster

Order of preference:

Common root cause across many tests (e.g. one primitive fix unlocked the entire UTF-8 cluster — 19 tests in one commit).
Named-param dispatcher pattern (:With/:By/:Of/:StartingAt unwraps) — usually 5–10 tests share the same fix shape.
Symbolic-position resolution (:First/:Last/:LastChar) — _ResolveSymPos in stzString already exists; extend its callers.
Per-method aliases — last resort, marginal returns.

Skip:

Standalone diagnostic tests (test_checker.ring, test_core.ring) that load the engine themselves — pre-existing, not part of the grind.
C27 syntax errors in test files — usually pre-existing.

3. Survey the cluster

# for the chosen error type, find call sites
for f in $(grep -l "^#ERR Error (R??)" $DIR/*.ring | head -10); do
    n=$(ring "$f" 2>&1 | grep "in file $(basename $f)" | grep -oE "line [0-9]+" | head -1 | awk '{print $2}')
    echo "$f L$n: $(sed -n "${n}p" "$f")"
done

Note: run from inside $DIR so relative ../../stzBase.ring resolves.

4. Apply the fix

Edit the offending method(s) in libraries/stzlib/base/string/stzString.ring (or stzList, stzStringFunc, etc.).

Before saving the final edit, mentally check:

Is the new method name unique case-insensitively? (Ring quirk)
Does the signature break any existing caller? grep -rn "\.MethodName(" libraries/stzlib/base/
Are operator precedences right? _aRes_ + (x + y - 1) not _aRes_ + x + y - 1

5. Syntax-check immediately

echo 'load "libraries/stzlib/base/string/stzString.ring"' > /tmp/loadcheck.ring
ring /tmp/loadcheck.ring 2>&1 | grep -iE "stzString|error " | head -3

If C22 fires, find the dup with:

grep -niE "^\s*def\s+<methodname>\s*\(" libraries/stzlib/base/string/stzString.ring

6. Spot-verify

cd libraries/stzlib/base/test/$TOPIC
for t in <list of cluster tests>; do
    echo "$t : $(ring "$t" 2>&1 | grep -E 'STOPPED|Error|panic' | head -1 | head -c 90)"
done

STOPPED! = pass. Anything else = still failing.

7. Commit + push both remotes

git add libraries/stzlib/base/string/stzString.ring  # be specific, not -A
git commit -m "string: <one-line cluster description>

<body with what was fixed and any caveats>

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>"

git push origin main
git push codeberg main  # ok to fail; codeberg auth expires

8. Re-annotate in background

python _annotate_test_errors.py $TOPIC
# ~25 minutes; run with run_in_background:true and continue working

9. Confirm strict decrease

When the background task completes, check the totals tail and the live count. If ERR went UP, something I introduced is wrong — search for case-insensitive duplicate defs first.

Iteration cadence

One cluster per commit. Don't batch multiple unrelated clusters into a single commit — if one regresses, the others are stuck with it. Small commits = trivial revert.

Reporting back

After each iteration, give the user a one-line update:

Cluster X (Y tests) — fixed Z. ERR floor: A → B. Annotate running in background.

Don't propose options for the next cluster — pick the highest-leverage one and start. The user can interrupt.

Constraints carried from CLAUDE.md

Use Softanza engine helpers, not Ring's substr byte-ops.
All test loops must be compile-time bounded.
No non-ASCII in console output.
Never use for X in list iterator form.
Hoist ring_len() out of for loops.
Push to both origin and codeberg.
Never create PRs; merge directly to main.

name	grind-err
description	Systematically reduce stzlib

Grind ERR

When to invoke

"grind string", "grind char", "grind "
"continue the grind", "keep going on path 1"
"reduce ERR count", "burn down failures"
After landing a fix, when the user says "continue"

The loop (one iteration)

1. Baseline

TOPIC=string  # or char, dataset, etc.
DIR=libraries/stzlib/base/test/$TOPIC

# live count of files carrying #ERR headers
grep -l '^#ERR' $DIR/*.ring | wc -l

# error-type distribution (the leverage map)
grep -h '^#ERR' $DIR/*.ring | sort | uniq -c | sort -rn | head -10

2. Pick the highest-leverage cluster

Order of preference:

Common root cause across many tests (e.g. one primitive fix unlocked the entire UTF-8 cluster — 19 tests in one commit).
Named-param dispatcher pattern (:With/:By/:Of/:StartingAt unwraps) — usually 5–10 tests share the same fix shape.
Symbolic-position resolution (:First/:Last/:LastChar) — _ResolveSymPos in stzString already exists; extend its callers.
Per-method aliases — last resort, marginal returns.

Skip:

Standalone diagnostic tests (test_checker.ring, test_core.ring) that load the engine themselves — pre-existing, not part of the grind.
C27 syntax errors in test files — usually pre-existing.

3. Survey the cluster

# for the chosen error type, find call sites
for f in $(grep -l "^#ERR Error (R??)" $DIR/*.ring | head -10); do
    n=$(ring "$f" 2>&1 | grep "in file $(basename $f)" | grep -oE "line [0-9]+" | head -1 | awk '{print $2}')
    echo "$f L$n: $(sed -n "${n}p" "$f")"
done

Note: run from inside $DIR so relative ../../stzBase.ring resolves.

4. Apply the fix

Edit the offending method(s) in libraries/stzlib/base/string/stzString.ring (or stzList, stzStringFunc, etc.).

Before saving the final edit, mentally check:

Is the new method name unique case-insensitively? (Ring quirk)
Does the signature break any existing caller? grep -rn "\.MethodName(" libraries/stzlib/base/
Are operator precedences right? _aRes_ + (x + y - 1) not _aRes_ + x + y - 1

5. Syntax-check immediately

echo 'load "libraries/stzlib/base/string/stzString.ring"' > /tmp/loadcheck.ring
ring /tmp/loadcheck.ring 2>&1 | grep -iE "stzString|error " | head -3

If C22 fires, find the dup with:

grep -niE "^\s*def\s+<methodname>\s*\(" libraries/stzlib/base/string/stzString.ring

6. Spot-verify

cd libraries/stzlib/base/test/$TOPIC
for t in <list of cluster tests>; do
    echo "$t : $(ring "$t" 2>&1 | grep -E 'STOPPED|Error|panic' | head -1 | head -c 90)"
done

STOPPED! = pass. Anything else = still failing.

7. Commit + push both remotes

git add libraries/stzlib/base/string/stzString.ring  # be specific, not -A
git commit -m "string: <one-line cluster description>

<body with what was fixed and any caveats>

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>"

git push origin main
git push codeberg main  # ok to fail; codeberg auth expires

8. Re-annotate in background

python _annotate_test_errors.py $TOPIC
# ~25 minutes; run with run_in_background:true and continue working

9. Confirm strict decrease

When the background task completes, check the totals tail and the live count. If ERR went UP, something I introduced is wrong — search for case-insensitive duplicate defs first.

Iteration cadence

One cluster per commit. Don't batch multiple unrelated clusters into a single commit — if one regresses, the others are stuck with it. Small commits = trivial revert.

Reporting back

After each iteration, give the user a one-line update:

Cluster X (Y tests) — fixed Z. ERR floor: A → B. Annotate running in background.

Don't propose options for the next cluster — pick the highest-leverage one and start. The user can interrupt.

Constraints carried from CLAUDE.md

Use Softanza engine helpers, not Ring's substr byte-ops.
All test loops must be compile-time bounded.
No non-ASCII in console output.
Never use for X in list iterator form.
Hoist ring_len() out of for loops.
Push to both origin and codeberg.
Never create PRs; merge directly to main.

grind-err

Grind ERR

When to invoke

The loop (one iteration)

1. Baseline

2. Pick the highest-leverage cluster

3. Survey the cluster

4. Apply the fix

5. Syntax-check immediately

6. Spot-verify

7. Commit + push both remotes

8. Re-annotate in background

9. Confirm strict decrease

Iteration cadence

Reporting back

Constraints carried from CLAUDE.md

Grind ERR

When to invoke

The loop (one iteration)

1. Baseline

2. Pick the highest-leverage cluster

3. Survey the cluster

4. Apply the fix

5. Syntax-check immediately

6. Spot-verify

7. Commit + push both remotes

8. Re-annotate in background

9. Confirm strict decrease

Iteration cadence

Reporting back

Constraints carried from CLAUDE.md