parseDates was using "02.01.2006" / "01/02/2006" which require
zero-padded fields. The Czech attendance sheet headers contain dates
like "1.6.2026", "23.3.2026", "6.4.2026" — Go silently dropped those
columns while Python's strptime accepted them. Effect was a missing
2026-06 month on /api/juniors plus undercounted attendance in any month
with single-digit columns; surfaced via make parity.
Use the unpadded reference forms "2.1.2006" / "1/2/2006" instead — Go's
time.Parse accepts both padded and unpadded inputs against them.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
/api/version returns each binary's own tag/commit/build_date, which
differs by design between independently built backends. Diffing it
always produces a false positive. Drop it from allRoutes; the route
remains reachable via `make parity ARGS="-route /api/version"`.
Also remove the vestigial `build_meta` allowlist entry (Python returns
the build dict as the top-level response body, not nested under
build_meta, so the scrubber never matched anything).
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
The transaction-row parser in services/membership/sources.go used a
helper (`getVal`) that did `fmt.Sprint(row[i])` before passing to
`matching.FormatDate`. The Sheets API returns date-formatted cells
as `float64` (Sheets serial-day numbers); pre-stringifying defeated
`FormatDate`'s `case float64:` dispatch, so values like 46147 leaked
through unchanged as the string "46147" instead of being converted
to "2026-05-05".
Surfaced by `make parity` (M5.4) — every `transactions[].date` on
/api/adults and /api/juniors differed between Python and Go. Python
side passes the raw value through directly (`isinstance(val, (int,
float))` in scripts/match_payments.py format_date), so it was always
correct.
Added a `getRaw` helper that returns row[i] without stringifying;
only the date column needs it. Extended TestLoadTransactions with
a numeric-serial-day row to lock in the regression.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Adds cmd/parity/main.go: a standalone Go binary that GETs
/api/version, /api/adults, /api/juniors, /api/payments from both
the Python (:5001) and Go (:8080) backends, scrubs an allowlist
(render_time.total, build_meta), and prints cmp.Diff for any
remaining differences. Exits 0 on full match, 1 on diffs, 2 on
fetch/parse errors — CI-friendly for M7.2.
- go/cmd/parity/main.go: flags (-py, -go, -route, -timeout), fetch
helper, allowlist scrubber (dotted-path aware), exit-code logic.
- go/cmd/parity/scrub_test.go: 4 unit tests for the scrubber.
- go/go.mod: promote github.com/google/go-cmp to direct dep.
- Makefile: parity target + help entry.
- Progress tracker: M5.4 ticked; milestone updated to M5 complete.
- Plan archived to docs/plans/2026-05-07-2254-m5-4-parity-binary.md.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Previously both backends defaulted to `CacheDir=tmp` and used the
same cache keys (`attendance_regular`, `attendance_juniors`,
`payments_transactions`, `exceptions_dict`) but stored different
shapes: Python caches post-processed view-model tuples
(e.g. `(members, sorted_months)`), Go caches raw sheet rows.
Whichever backend wrote last poisoned the cache for the other,
producing `ValueError: too many values to unpack (expected 2,
got 68)` on Python's /adults after the Go side populated the
file with 68 raw CSV rows.
This breaks the M5.4 `make parity` workflow that requires both
backends running side-by-side.
Fix: change Go's default to `tmp/go` so the two cache trees
never overlap. `CACHE_DIR` env var override still works.
`os.MkdirAll` already handles creating the new subdirectory on
first write.
Recovery for users with poisoned `tmp/`: hit /flush-cache on
the Python side once after pulling, then restart the Go server.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Mirror fuj infer's read-only mode: SyncOpts.DryRun skips WriteHeader,
AppendValues, and SortByDateColumn, printing one "Dry run: would …"
line per planned operation instead. ID-dedup still runs so the output
reflects exactly what the next real sync would write.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Closes M3.1–M3.6. Parity safety net proving Go output matches Python
for every ported pure-domain function (M2.1–M2.9) and reconcile (M2.10).
Capture pipeline:
- scripts/capture_fixtures.py: calls each Python function with seeded
inputs, emits JSON fixtures to stdout (never writes files directly).
- scripts/scrub_fixtures.py: deterministic PII scrubber — SHA-256
pseudonyms for member names, digit-preserving hashes for VS/account/
bank_id, name-sweep in message text. Idempotent; no salt.
- scripts/_fixture_seeds.py: handcrafted seeds for all 11 functions;
synthetic names throughout (no real roster members).
- scripts/capture_all_fixtures.sh: convenience wrapper for full corpus
regeneration outside of make.
Fixture corpus (98 files, all PII-free):
- go/tests/fixtures/pure/<func>/<case>.json — 10 function directories.
- go/tests/fixtures/reconcile/<NN>_<case>.json — 10 branch-coverage
cases: greedy, overpayment credit, proportional remainder, even-split,
out-of-window, exception override, other: purpose, junior ?, multi-
person+month fan-out, unmatched.
Go parity tests (//go:build parity):
- go/tests/parity/parityio.go: generic LoadDir/RunAll helpers + typed
In/Out struct pairs for all 10 pure functions; Envelope decoder for
int/float/none disambiguation.
- 10 pure-function test packages + bespoke reconcile test with per-cell
float tolerance (math.Abs <= 0.01 for `paid` values).
Makefile: go-parity, go-test-all, capture-fixtures targets.
go/tests/fixtures/README.md: refresh workflow + PII audit guide.
Gate: make go-test green, make go-parity green (11/11 packages),
make go-lint clean (parity tag), make go-build clean.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Add internal/services/membership package: AttendanceLoader,
TransactionLoader, ExceptionLoader interfaces + NewStubSources stub
(returns ErrIOPending until M4 lands real Sheets loaders).
FeesReport and ReconcileReport orchestrate domain/fees + domain/reconcile
and write fixed-width text reports matching Python calculate_fees.py and
match_payments.py print_report output. 13 unit tests cover all formatter
branches and orchestration wiring via fake loaders.
cmd/fuj/main.go: fees and reconcile subcommands now dispatch; sync/infer
retain the [M4] placeholder.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
infer_payments was building member_names from get_members_with_fees()
(adults sheet only). Junior-only members were invisible to the matcher,
so a payment message containing an exact junior name would produce a
fuzzy review match against a different adult sharing the same first name.
Fix: union the adult and junior rosters (deduped via canonical_member_key)
so all members are candidates. The existing exact-name short-circuit in
match_members then handles precedence correctly.
Two regression tests added for the Jáchym Kubík / Jáchym Hrušák case.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
New go/internal/domain/matching package porting three helpers from
scripts/match_payments.py:
- BuildNameVariants: normalized ASCII variants from a member name (nickname
in parens, last/first split, len<3 filtered); variants[0] is always the
full base name — MatchMembers relies on this invariant.
- MatchMembers: auto/review confidence matching with an exact-name
short-circuit pass that prevents nickname substrings (tov) from firing
inside longer surnames (ottova); common-surname filter for review tier.
- FormatDate: nil/empty/""/serial int/float64 (since 1899-12-30, fractional
days supported)/YYYY-MM-DD passthrough/garbage → never errors.
- InferTransactionDetails: composes BuildNameVariants+MatchMembers+
ParseMonthReferences; falls back to sender-only member match and
date-derived month when text carries no signal.
21 table-driven tests; all expected values verified against live Python
on 2026-05-06. go-build, go-test, go-lint all clean.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
SHA-256 dedup hash from sync_fio_to_sheets.py generate_sync_id.
Key subtlety: Python str(float) emits "500.0" for whole-valued floats
and switches to scientific notation at |f|>=1e16 or |f|<1e-4 —
replicated via formatAmount using 'f'/'e' format selection.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Pin Sep 2025 – Feb 2026 at 750 CZK in ADULT_FEE_MONTHLY_RATE so
historical billing is unchanged; new default 700 CZK applies to
2026-04 onward. March 2026 stays at 350.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
The previous calculation derived balance from total_balance (which includes
current/future-month activity and out-of-window credits) plus a one-sided
debt-only adjustment. Current-month surplus leaked through, making the balance
appear less negative than actual past-month debt (e.g. Mauric Daniel -1250 vs
correct -1750). Pay-All is now max(0, -balance) so the two values share a
single source and cannot disagree.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
reconcile() previously split a multi-month payment evenly across months,
which falsely flagged months as underpaid when their expected fees
differed (e.g. 1250 CZK for 02+03+04 2026 with rates 750/350/150 was
shown as 416/month with two months red).
The allocation now runs per matched member: greedy when the share covers
the total expected (each month gets its expected fee, surplus -> credit),
proportional by expected fee otherwise. Out-of-window months keep the
previous even-split-to-credit behavior. 6 new test cases.
Also adds CHANGELOG.md and a changelog convention in CLAUDE.md.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
