Two hot-path fixes for DBF reads surfaced by the bulk-bench profile.
1. parseNumericField decimal path — was 23% of flat CPU on BULK_CTE.
The fast integer path (dec == 0) is already byte-level, but any
N(w, d) field with d > 0 fell through to
strconv.ParseFloat(string(raw[start:end]), 64)
allocating per-row. A 10k-row CTE insert ran this 200k+ times.
Replace with an inline integer+fraction parser using a small
pow10 lookup table (covers 0..19 decimal places). Unexpected
characters still fall back to strconv for correctness.
Result:
BULK_CTE_10k_20iter 187 → 83 ms (2.25x)
BULK_SUBQ_10k_20iter 102 → 22 ms (4.6x)
2. DBFArea.RecCount in shared mode was doing Seek(0, 2) on every
call. SqlScan calls it once per query for its result-array
pre-allocation (~0.2 ms × 1000 queries = 0.2s of CPU on the
bench). Cache the count per-area, keyed by a process-wide
generation counter. Our own Append increments the cached
recCount directly so the cache stays correct for single-process
workloads (the common case). Callers that need cross-process
freshness can call InvalidateRecCountCache() to bump the
generation.
SQL bench: modest 1-3 ms drops on B1/B2/B3/B6/B7.
Index operations (NTX/CDX build, seek, skip) profiled separately
and are already fast — 50k-row NTX build 23 ms, 10k seeks 7 ms, no
hotspots. Left untouched.
FiveSql2 43/43, Harbour compat 56/56, Go test ALL PASS.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
parseNumericField was allocating on every call — `string(raw)` to
convert the record-buffer slice to a string, plus the implicit
allocation from TrimSpace's return value. For a 50k-row scan reading
two numeric fields, that's 100k+ small string allocations per scan,
all of which promptly became garbage.
Rewritten to walk the raw byte slice directly:
- Find the trimmed range by byte indexing (no alloc).
- Parse integer-typed fields (dec == 0) digit-by-digit into int64.
- Only fall back to strconv.ParseFloat + string allocation for
genuinely fractional data (dec > 0 or embedded `.`).
This also lifts the raw RDD baseline in our bench (6.8ms → 6.2ms)
because FieldGet hits this same parser. Every scan path benefits,
not just the FiveSql2 hot loop.
Measured (50k rows, 3-run steady state):
Before After
No WHERE 10.0ms 9.1ms
Numeric WHERE 7.8ms 6.9ms ← now 1.11x raw
String WHERE 7.9ms (see next commit)
Raw RDD baseline 6.8ms 6.2ms ← also faster
Validation:
- hbrdd/dbf tests PASS (including integer/float field roundtrips)
- FiveSql2 43/43
- Harbour compat 51/51
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Replaces the `return NIL` stubs with real implementations that read
from the current workarea. Covers the info codes actually used by
downstream code (FiveSql2 TSqlIndex, standalone callers):
DBINFO:
DBI_ISDBF, DBI_CANPUTREC, DBI_FULLPATH, DBI_TABLEEXT, DBI_MEMOEXT,
DBI_SHARED, DBI_ISREADONLY, DBI_GETRECSIZE, DBI_DBVERSION,
DBI_RDDVERSION, DBI_BOF, DBI_EOF, DBI_FOUND, DBI_FCOUNT, DBI_ALIAS,
DBI_POSITIONED
DBORDERINFO:
DBOI_EXPRESSION, DBOI_NAME, DBOI_NUMBER, DBOI_POSITION,
DBOI_ORDERCOUNT, DBOI_KEYCOUNT, DBOI_KEYCOUNTRAW
Unknown info codes still return NIL (Harbour's forgiving fallback).
New accessors on DBFArea (FullPath, IsShared, IsReadOnly) expose the
private filePath/shared/readOnly fields to the hbrtl layer without
plumbing them through the generic Area interface.
Unblocks TSqlIndex:FindExclusive's original DBI_FULLPATH/DBI_SHARED
scan — though the short-circuit there stays in place for now since
it's a correctness workaround that no longer masks a crash thanks
to the recent gengo PushMemvar fallback.
Validation:
- FiveSql2 43/43 (0 warnings)
- Harbour compat 51/51
- go test ./... ALL PASS
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Two SQLite-style optimizations for RDD and SQL workloads:
1. FieldPos() O(1) column binding cache
Before: FieldPos(name) linear scan — O(n) per call with string
comparison. In SQL engines that call FieldPos per row per
column, this is hundreds of thousands of calls.
After: DBFArea builds a map[UPPER(name)]→pos on first lookup.
All subsequent lookups are O(1) hash. SQLite calls this
"column affinity binding" — positions resolved at prepare,
not per row.
Implementation:
- hbrdd/dbf/dbf.go: DBFArea.FieldPosCache(name) method
- hbrtl/procinfo.go: FieldPos RTL uses fieldPosCacher interface
- Lazy init: only pays for tables that get queried
2. hbrdd import auto-detection for function-call style PRGs
Before: compiler only added hbrdd import when PRG used xBase commands
(USE, SKIP, INDEX...). Pure function-call style like
`dbUseArea(.T.,,"t")`, `FieldPut(1, val)` was missed —
generated Go failed to compile ("undefined: hbrdd").
After: scanStmtsForXBase walks ExprStmt bodies too, detecting
CallExpr to any of the ~40 xBase RTL function names.
FIELD->NAME alias expressions also trigger the import.
Resolves: small PRGs that use only dbUseArea/FieldGet/FieldPut.
Benchmark notes (50k records):
Raw RDD scan: 7 ms (baseline)
FiveSql2 SELECT WHERE: 157 ms (unchanged — bottleneck is
not FieldPos, it's PRG-level
expression tree walk per row)
compat_harbour 51/51: PASS
FiveSql2 43/43: 100%
The FieldPos cache helps heavy field-name-based code paths but the
primary FiveSql2 bottleneck is the PRG interpreter walking expression
ASTs per row (needs bytecode compilation to close the gap).
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Eliminate MacroEval overhead for INDEX ON with UDF/complex expressions.
Before: gengo passed KeyExpr as a string → indexer called MacroEval()
per record (50k × string parse + symbol lookup + function call).
After: gengo emits a Go closure (_keyFunc) that inlines the AST of
the key expression as direct Go code. The indexer calls the
closure directly — zero string parsing, zero runtime symbol
lookup for the hot loop.
Three code paths in the closure, depending on expression type:
1. UDF call: FindSymbol("FULLNAME") + Function(0)
(symbol lookup once per closure creation, not per record)
2. Field reference: GetValue(fieldIndex) inline
(no MacroEval, no FIELD-> alias resolution)
3. UPPER/LOWER(expr): strings.ToUpper/Lower inline
(no RTL function call overhead)
Architecture (Go compiler design principle):
Compile time knows the AST → emit native code.
Don't serialize to string → re-parse at runtime 50k times.
Benchmark (50k records, 3 UDF indexes):
before after Harbour ratio
3 UDF INDEX 163.0ms 60.0ms 55.0ms Five/HB = 1.09x
SEEK 10k 7.6ms 7.6ms 14.0ms Five 1.8x faster
SCAN 50k 3.4ms 3.4ms 4.0ms Five 15% faster
TOTAL 233.0ms 130.0ms 147.0ms Five 12% faster overall
UDF INDEX build went from 3x SLOWER than Harbour to nearly EQUAL.
SEEK/SCAN remain faster than Harbour (mmap + NTX optimizations).
Changes:
hbrdd/driver.go KeyFunc field in OrderCreateParams
hbrdd/dbf/indexer.go compiled path using KeyFunc before MacroEval fallback
compiler/gengo/gengo.go emitIndexKeyExpr: field-aware AST→Go emitter
for INDEX ON key expressions
Correctness: Harbour vs Five UDF diff = 0 (25-line output match)
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
All 3 remaining known constraints resolved. CLAUDE.md now shows zero.
1. CDX compound index WRITE support (was read-only)
New file: hbrdd/cdx/build.go (~400 LOC)
- CreateOrAddTag() builds Harbour-compatible CDX files
- Bit-packed leaf pages (RecBits/DupBits/TrlBits compression)
- Interior nodes with big-endian RecNo/ChildPage
- Compound root directory (structural B-tree of tag names)
- Append-safe: preserves existing tags when adding new ones
- Linked leaf pages (LeftPtr/RightPtr for sequential scan)
Pipeline: INDEX ON expr TAG tagname TO file
- ast.IndexCmd gains TagName field
- Parser captures TAG name (was discarded)
- gengo passes TagName to OrderCreateParams
- indexer.go routes to cdx.CreateOrAddTag when TAG specified
Verified: 3 tags (BYNAME/BYCITY/BYAGE), OrdSetFocus by name,
SEEK, GoTop/GoBottom, close+reopen with SET INDEX TO
2. {||} empty code block parsing in function arguments
Parser's parseArrayOrBlock() called parseExpr() unconditionally
after closing |, failing when body was empty ({||}).
Fix: check for RBRACE after closing | and emit NIL literal body.
{=>} empty hash already worked.
3. Semicolon IF...ENDIF — already worked (removed from constraints)
Tests:
go test ./... 14 packages ALL PASS
FiveSql2 43/43 100%
compat_harbour 51/51
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Replace the no-op Windows lock stub with actual kernel32 LockFileEx /
UnlockFileEx calls via syscall.LazyDLL (zero external dependency).
- LOCKFILE_EXCLUSIVE_LOCK | LOCKFILE_FAIL_IMMEDIATELY for non-blocking
semantics matching Clipper FLOCK() → .F.
- Same lock region layout as POSIX: header region for FLOCK, record
offsets for DBRLOCK — compatible across platforms
- Handles returned as syscall.Handle from os.File.Fd()
Note: full Windows cross-compile still blocked by unrelated issues
(mmap in cdx/ntx, termios in debugcli.go). The lock code itself
compiles cleanly with //go:build windows.
Also updates gap-analysis.md to reflect Windows lock status.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Replaces the FLOCK/DBRLOCK/DBRUNLOCK no-op stubs with actual
fcntl(F_SETLK) byte-range advisory locks, matching Harbour's
hb_fsLockLarge implementation.
Before: rtlDbRLock always returned .T. regardless of contention.
Multi-process writers could silently corrupt records.
After: Non-blocking POSIX byte-range locks per file descriptor.
Cross-process exclusion verified by a subprocess-spawning
Go test that witnesses BUSY vs OK transitions.
New files:
hbrdd/dbf/locks_posix.go fcntl F_WRLCK/F_UNLCK wrappers
hbrdd/dbf/locks_windows.go stub (TODO: LockFileEx)
hbrdd/dbf/lock_multi_test.go cross-process verification
docs/gap-analysis.md honest Harbour parity assessment
Modified:
hbrdd/dbf/dbf.go
- DBFArea gains fileLocked bool + lockedRecs map
- Close() calls releaseAllLocks() before dropping the fd
hbrtl/database.go
- rtlDbRLock / rtlDbRUnlock now delegate to DBFArea.LockRecord /
UnlockRecord instead of returning fixed .T./NIL
- New rtlFLock / rtlDbUnlock for FLOCK() / DBUNLOCK()
hbrtl/register.go
- FLOCK and DBUNLOCK symbols registered (were missing entirely)
compiler/analyzer/analyzer.go
- FLOCK / DBUNLOCK added to RTL known-function set
Lock region layout (non-overlapping on purpose):
FLOCK region [0, HeaderLen+1)
Record N region [RecordOffset(N), RecordLen)
So a workarea can hold FLOCK and multiple DBRLOCK simultaneously
on the same fd without conflict.
Design rationale (captured in locks_posix.go header):
* POSIX fcntl, not flock(2) — byte-range + NFS-safe
* Non-blocking F_SETLK — matches Clipper FLOCK() → .F. semantics
* Released explicitly on Close to avoid workarea-sharing races
* Windows falls back to no-op (TODO: LockFileEx)
Verification:
go test ./hbrdd/dbf/ -run TestFLockBlocksAcrossProcesses PASS
go test ./hbrdd/dbf/ -run TestRLockBlocksAcrossProcesses PASS
go test ./... ALL PASS
FiveSql2 43/43 100%
compat_harbour 51/51 100%
The gap-analysis doc (docs/gap-analysis.md) is a running inventory
of what works vs what's still missing vs Harbour 3.2, written for
users evaluating Five for production — not a sales pitch.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Benchmark (50k records, 4 indexes on Apple M-series):
before after Δ
INDEX 53.7ms 33.3ms -38% (now 10% faster than Harbour 37.3ms)
TOTAL 156.2ms 133.0ms -15%
Fixes:
1. sort.Slice(reflection) → concrete sort.Interface
Benchmarked in isolation on 200k KeyRecords:
sort.Slice(closure): 50.0ms
sort.Sort(interface): 30.4ms (40% faster, no reflection)
- indexer.go: add keyRecordAsc/Desc concrete types
- Branch hoist descending check out of Less()
2. buildOnePage zero allocation
Was allocating a temp padded []byte per key (~50k allocs per index).
Now writes padded key directly into the page buffer via padCopy.
3. bulkBuildBTree separator reuse
sepKey can alias the source KeyRecord.Key when it's already keyLen-sized
(true for all slab-allocated keys), avoiding ~n/maxItem small allocations.
Pre-size the children slice.
4. Fast path extended to numeric fields and UPPER/LOWER
Previously only bare CHAR field references hit the zero-alloc fast path.
Now:
- Numeric fields (N/F type) copy DBF bytes directly
(same-length ASCII compare matches numeric order for non-negatives)
- UPPER(field) / LOWER(field) wrappers on CHAR fields apply ASCII
case folding inline during byte copy
Per-index timing on the micro benchmark:
before after
NAME 7.7ms 7.5ms (fast path, unchanged)
CITY 6.0ms 6.2ms (fast path, unchanged)
AGE 14.1ms 7.1ms -50% (was slow path)
UPPER(NM) 17.0ms 7.9ms -54% (was slow path)
5. Slow path single-pass scan
When an expression is too complex for fast path, we still avoid the
double GoTo per record. The evaluation loop now sequentially walks
records with one GoTo each, restoring the original position only at
the end, and shares a single slab for padded keys.
Also fixes a hbrt bug surfaced while writing the benchmark:
6. Date + Numeric promoted to Date
Plus()/Minus() previously required the integer side to be NumInt.
Modulus returns a promoted type, so `SToD("...") + (i % 365)` panicked.
Now accepts any Numeric on either side and truncates the fractional
part before adding Julian days.
- hbrt/ops_arith.go: Date±Numeric (was Date±NumInt only)
Tests:
go test ./... — ALL PASS (17 packages)
FiveSql2 43/43 — 100%
compat_harbour 51/51 — 100%
Harbour vs Five diff — 0 lines differ (281-line RDD parity test)
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
From senior Go developer review:
C7 CRITICAL: pagePool data race (ntx.go)
- Moved global pagePool[8] + pagePoolIdx into per-Index struct
- Eliminates race condition across goroutines using separate indexes
C8 CRITICAL: Page.data dangling pointer after remap (ntx.go)
- remapFile() now clears pagePool data slices (pointed into old mmap)
- Prevents segfault from stale mmap references
C4 HIGH: pop() bounds check restored (thread.go)
- Removed performance optimization that eliminated underflow detection
- Stack underflow now produces clear error instead of index -1 panic
C1 HIGH: intExpLen overflow on MinInt64 (value.go)
- Added special case: MinInt64 returns 20 (length of -9223372036854775808)
- Prevents -v overflow in negation
C11 CRITICAL: GoTo ReadAt error handling (dbf.go)
- ReadAt failure now returns error and sets EOF
- Previously silently used stale record buffer (data corruption risk)
C14 HIGH: LEN() inline missing Hash case (gengo.go)
- Added _v.IsHash() → len(Keys) branch
C15 HIGH: EMPTY() inline missing Date case (gengo.go)
- Added _v.IsDate() && _v.AsJulian() == 0 check
82/82 stress PASS. 14 packages ALL PASS.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
NTX Bulk Build (build.go — ported from rddfive/ntx_engine.c):
- pageBuffer: dynamic memory buffer for all pages
- Phase 1: Build leaf pages in sequential memory (zero disk I/O)
- Phase 2: Build interior levels from cached leaf data (zero I/O)
- Separator promotion: remove last key from leaf only (not interior)
- Single bulk WriteAt for all pages at end
- INDEX ON 10K: 34ms → 5-8ms (4-6x improvement)
NTX Seek (ntx.go):
- Always descend to leaf on match (find first occurrence)
- fStop flag tracks path match, verified at leaf
APPEND Buffering (dbf.go):
- Append marks dirty without immediate disk write
- flushRecord writes record data only (no header/EOF per record)
- Close/Flush writes EOF marker + header once
Results: 14 packages ALL PASS, 82/82 stress test
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
NTX Seek (ntx.go):
- Always descend to leaf even on internal match (Harbour behavior)
Prevents SEEK returning internal separator instead of first leaf entry
Fixes duplicate key SEEK (NYC=9→10, Paris=8→10)
- fStop flag tracks path match, verified at leaf with key comparison
- Handle fStop at page end: ascend via nextKey to find actual match
SET DELETED + SEEK (indexer.go):
- When SEEK finds a deleted record with SET DELETED ON:
Skip forward through matching deleted records
If all matching records deleted → return not found (EOF)
Fixes H04: deleted record now correctly returns .F.
BOF (indexer.go + dbf.go):
- Set a.FBof AFTER a.GoTo returns (GoTo resets FBof=false at line 393)
- Fixes infinite loop in DO WHILE !BOF() ... SKIP -1
Results:
- Unit tests: 14 packages ALL PASS
- 77-item thorough test: 77/77 (100%)
- 82-item stress test: 82/82 (100%) — Harbour identical
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
1. SOFTSEEK: use idx.CurRecNo() for positioning (was checking recNo > 0)
- SEEK with SET SOFTSEEK ON now positions at next higher key
- SEEK command reads SET SOFTSEEK at runtime (was compile-time only)
- rtlDbSeek defaults to GetSetSoftSeek() when no explicit param
2. SET DELETED ON + INDEX: SkipIndexed skips deleted records
- GoTopIndexed: skip deleted record at top position
- SkipIndexed: inner loop continues past deleted records
3. Compound key (CITY+NAME): field name TrimSpace before lookup
- evalKeyExprInner: TrimSpace on fieldName after FIELD-> strip
- Fixed "CITY " != "CITY" mismatch from + operator splitting
4. SET INDEX TO filename: treated as string, not variable
- gengo uses exprToString for SET INDEX TO (was emitExpr)
- Prevents identifier being resolved as local variable
5. hasXBaseCommands: recursive scan into nested blocks
- BEGIN SEQUENCE, IF, FOR, DO WHILE, SWITCH bodies now scanned
- Fixes missing hbrdd import for DB commands inside blocks
Thorough test: 77 items (14 sections) covering exact/partial/soft seek,
SET DELETED, duplicate keys, numeric keys, compound keys, empty/single
table, state consistency, order switching, full traversal — all identical.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Core change:
- dbf.KeyEvalFunc: global callback set by gengo before OrderCreate
- evalKeyExprInner default case: calls KeyEvalFunc for unknown functions
- Final fallback: any unresolvable expression → KeyEvalFunc → MacroEval
- valueToKeyBytes: converts MacroEval result to index key bytes
- gengo: sets dbf.KeyEvalFunc = t.MacroEval before OrderCreate, clears after
Examples that now work:
INDEX ON MyFunc(FIELD->NAME) TO idx // UDF in key expression
INDEX ON CityKey(FIELD->CITY, NAME) TO idx // multi-param UDF
INDEX ON Left(MyFunc(NAME), 15) TO idx // nested built-in + UDF
Also fixed:
- SET ORDER TO n: int→string via hbrt.NtoS (was empty string)
- CDX compound leaf decoder: proper bit-packed tag name extraction
- CDX compound recNo = direct byte offset (not page number)
All existing tests pass, NTX 47/47 + CDX 20/20 Harbour compat maintained.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
CDX Integration:
- IndexEngine interface: common for NTX Index and CDX Tag
- OrderListAdd: auto-detects .cdx/.ntx extension, opens CDX tags
- decodeCompoundLeaf: proper bit-packed tag directory decoding
(was stub falling through to scanCompoundLeaves with wrong names)
- CDX Tag: added KeyLen(), KeyExpr(), ForExpr(), IsDescending(), Close()
- CDX compound recNo = direct byte offset (not page number)
ORDSCOPE:
- SetScope/ClearScope/SetScopeTop/SetScopeBottom on DBFArea
- GoTopIndexed: seeks to scopeTop, validates within scopeBottom
- GoBottomIndexed: seeks to scopeBottom boundary
- SkipIndexed: stops at scope boundaries (top and bottom)
- OrdScope RTL function registered (nScope: 0=TOP, 1=BOTTOM)
- scopeKeyFromValue: converts Value to padded key bytes
Index Order Management:
- OrderListFocus: handles numeric order ("2" → order 2)
- SET ORDER TO n: gengo emits hbrt.NtoS for int-to-string conversion
- IndexOrd/OrdCount/OrdName/OrdKey: real implementations (were stubs)
- OrderCount/CurrentOrder/OrderName/OrderKeyExpr accessors on DBFArea
- ClearScope on order switch (prevents stale scope)
Cross-read test: Harbour-created CDX → Five reads, 20/20 items match:
NAME/CITY/ID seek, ORDSCOPE count, GoTop/GoBottom all identical
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
sort.Slice is unstable: equal keys had random record order.
Harbour NTX B-tree orders equal keys by ascending RecNo.
Added RecNo tiebreak to sort comparator.
Result: 47/47 (100%) Harbour compatibility on rdd_compat test.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Bug 1: FIELD->NAME in INDEX ON expression
- evalKeyExprInner: strip FIELD->/alias-> prefix before field lookup
- exprToString: handle AliasExpr (FIELD->NAME → "FIELD->NAME")
Bug 2: AsNumInt() on Double returned IEEE 754 raw bits
- Value.AsNumInt(): check tDouble and convert via Float64frombits
- Fixed array index crash when index is result of % modulo
Bug 3: PACK/ZAP crash with open indexes
- OrderListRebuild: fully implemented (was TODO stub)
Saves index info, closes all, sets idxState=nil, recreates
- OrderCreate: set current=-1 during key evaluation (natural GoTo)
- PACK/ZAP: save/restore idxState, rebuild after operation
- Register __DBPACK, __DBZAP, DBRECALL symbol aliases
Harbour vs Five: 45/47 match (96%), 2 diffs are duplicate-key sort order
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- skipFilter: skip deleted records in GoTop/GoBottom/Skip when SET DELETED ON
- hbrdd.IsSetDeleted callback: avoids circular import hbrdd→hbrtl
- Parser: capture ON/OFF for boolean SET commands (DELETED, EXACT, SOFTSEEK, etc.)
- Parser: capture TO expr for SET DATE/DECIMALS/EPOCH
- Gengo: emit proper t.Do() calls for 11 SET toggles + 3 value SETs
- stmtSet: was stub (skipToEOL), now calls parseSet()
- RTL: register 11 SET toggle functions (SETDELETED, SETEXACT, etc.)
- RTL: DBLOCATE/DBCONTINUE for sequential search
- RTL: DBSETFILTER/DBCLEARFILTER/DBFILTER
- PadL/PadR: support 3rd param fill character
- Area interface: added SetFound, SetLocate, LocateBlock, filter methods
- MemRDD: implements new Area interface methods
- Comprehensive PRG test: test_search.prg (7 test suites all pass)
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
