audit_decisions.md

Audit follow-up: decisions log (Sept 2026)

Decisions made while fixing the findings of the September 2026 full-project audit. Items marked (user) were chosen by Eric; items marked (auto) were made autonomously overnight under the standing instruction "go ahead with your recommendations and record them" — review these.

Task numbers (#N) refer to the audit task list.

Semantics / dialect

#	Decision	Who	Notes
40, 44	Target CoCo Color BASIC semantics for shared features, keeping the modern extensions (MOD, WHILE, DO, LOCAL/PRIVATE, labels).	user
4	RND(0) returns a fresh fraction in [0,1), as on CoCo (not MS BASIC-80's repeat-last). RND(x) for 0<x<1 behaves like RND(0). last_rnd removed.	user	Old test_rnd_semantics.py repeat-last tests rewritten.
4	Per-interpreter RNG (CoCoBasic.rng); RND(-n) and RANDOMIZE no longer reseed other sessions.	auto	Unambiguous bug fix.
5	INT is floor (INT(-3.5) = -4).	auto	True of every Microsoft-derived BASIC incl. Color BASIC. Two old tests asserted truncation; corrected.
10	Undimensioned arrays auto-dimension to 10 on first use.	user	Old tests that used UNDIM(999)=5 to force an error now expect BAD SUBSCRIPT. A consequence: an unknown name with parentheses, e.g. FOO(5), is an array (as on CoCo), not an "undefined function" error.
10	DIM A(0) legal; total elements capped at 1,000,000 (MAX_ARRAY_ELEMENTS, variables.py) → OUT OF MEMORY error.	auto	Cap value is a judgment call (server memory safety).
6	Empty DATA items (DATA 1,,3) read as 0 / "". READ into a numeric variable from non-numeric DATA → TYPE MISMATCH error (real CoCo says ?SN ERROR in the DATA line; the clearer message was preferred). READ into a string variable converts numbers to their text.	auto	split_args(keep_empty=True) added to the shared splitter rather than a new one.
9	VAL follows MS BASIC prefix parsing: spaces ignored, &H/&O accepted, stops at the first non-numeric char, overflow → error. Implemented as basic_number_prefix() in functions.py.	auto
12	STRING$ code outside 0–255, or an empty string argument → ILLEGAL FUNCTION CALL.	auto
84	CIRCLE ratio and arcs: CIRCLE(x,y),r,c,ratio,start,end. The x radius is r, the y radius r × ratio (0–4); start/end are fractions of a turn (0–1) clockwise from 3 o'clock, end < start wraps. With ratio 1 and no arc the midpoint circle is kept; otherwise a 64-segment polygon (arc_points / arcPoints, identical arithmetic in Python and JS). Out-of-range values are ILLEGAL FUNCTION CALL.	auto	The CoCo's own arc drawing isn't reproduced pixel-exactly.
99	PRINT items next to a string literal need no separator (PRINT "X="X, A$"!", "A""B"), printed as after ;, as in Color BASIC. Two non-string items still need one (PRINT 1 2 stays a syntax error: with BasiCoCo's spacing rules it can't mean the CoCo's crunched 12). PRINT# now parses its items with the same PRINT parser, replacing file_io's own tokenizer (the #76 note).	auto
118	An INPUT answer that can't be stored is a runtime error at the INPUT statement: OVERFLOW for a number too large (as VAL), BAD SUBSCRIPT for an index out of range. store_input_value keeps it as pending_input_error and continue_program_execution raises it, so ON ERROR can trap it and no front end can drop it (the web server and CLI ignored the return value). Non-numeric text still reads as 0 (VAL's rule), not ?REDO.	auto	Open question: re-ask (?REDO) on non-numeric input as Color BASIC does?
87	PEEK, POKE, VARPTR, USR (USR0–USR9) and EXEC are reserved and unsupported: each gives "ILLEGAL FUNCTION CALL: … is not supported in BasiCoCo" (ERR 5). There is no simulated 6809 memory map; that would be a separate feature.	auto
124	One colour model for PPOINT and the canvas: COLOR fg,bg sets the drawing colour and background (a new background clears the screen to it, as the client always did); PRESET and LINE…,PRESET draw the background colour (was 0 on the client, the PCLS colour on the server); PCLS alone and PMODE clear to the background. The server's pixel record now follows every one of these.	auto	PMODE clearing the screen mirrors the client; on a real CoCo PMODE only reinterprets video memory.
116	A missing start line for CHAIN "f",n or RUN n stops the program; ON ERROR doesn't trap it. RUN clears ON ERROR before starting, and a chained program that never started has no statement to RESUME to (the old handler's line number means nothing in it). Jumps inside a running program are trappable (#105).	auto
102	ERR uses Microsoft BASIC's error numbers (NF 1, SN 2, RG 3, OD 4, FC 5, OV 6, OM 7, UL 8, BS 9, /0 11, TM 13, LS 15, RESUME WITHOUT ERROR 20, WEND WITHOUT WHILE 30, bad file number 52, file not found 53, bad file mode 54, file already open 55, input past end 62). ERR/ERL/ON ERROR are Microsoft's design, so their numbers are the natural fit; the old table mixed schemes (division by zero was 99). Unrecognised errors are 0.	auto	Open question: CoCo 3's ERNO numbers errors differently (/0 = 10); switch if CoCo 3 compatibility matters more.
103	PRINT# writes items exactly as PRINT shows them (format_print_item(), ast_nodes.py): numbers keep the sign position and trailing space, so PRINT #1,5;6 writes 5 6. INPUT # ends a numeric item at a space (as well as a comma or line end), as in Microsoft BASIC, so that line reads back as 5 and 6. A comma in PRINT# still writes a literal , rather than padding to a print zone, so PRINT #1,A$,B$ reads back as two fields; address_book.bas relies on it.	auto	Open question: Color BASIC's own PRINT# comma behaviour was not verified.
86	Strings are not limited to 255 characters (Color BASIC gives ?LS ERROR). Kept as a modern extension: it breaks no working CoCo program.	auto	Open question: enforce 255 for fidelity?
97	RENUM takes Extended Color BASIC's order, RENUM new,start,increment (was new,increment,start); empty positions keep their defaults (RENUM ,,5). A single line may now be renumbered to the top (65530), which the old overflow check refused. Line numbers still go to 65535, not CoCo's 63999.	user delegated ("do what makes the most sense")	Two-argument calls with increment 10 (RENUM 100,10) mean the same in both orders on programs starting at line 10 or above.

Filesystem sandbox

#	Decision	Who	Notes
1, 3	Sandbox everywhere (web server and local basicoco.py). Every filename resolves inside ./programs (the writable root) via FileManager.resolve_writable() / find_readable(). Absolute paths, ~, drive letters and .. are refused; symlinks resolving outside are rejected.	user
1	The project's bundled programs/ is a read-only fallback for LOAD / MERGE / CHAIN / OPEN "I". KILL, SAVE and OPEN "O"/"A" only touch the writable root — so KILL can no longer delete a bundled program when the server runs from another directory.	auto
3	CD is a per-session virtual directory inside programs/ (FileManager.subdir); os.chdir is never called. CD "/" and CD "~" go to the programs root; CD ".." stops at the root; CD alone prints CURRENT DIRECTORY: PROGRAMS/....	user (auto: "~" and "/" both mean root)	Old CD tests that asserted a real chdir were rewritten.
1	Paths are relative to programs/: SAVE "sub/pathtest" writes programs/sub/pathtest.bas (a test that saved "programs/pathtest" was updated).	auto
2	KILL records the target server-side (FileManager.pending_kill); the confirmation carries only Y/N and any client-sent filename is ignored. The input_request no longer leaks the absolute path. KILL confirmation now also works in the standalone basicoco.py CLI (it was never wired up).	auto
11	LOAD / SAVE / MERGE / CHAIN / KILL / CD evaluate string-expression filenames (SAVE F$), like OPEN already did.	auto
7	OPEN reports every OSError (e.g. a directory) as a BASIC FILE ERROR; INPUT#/LINE INPUT# report BAD SUBSCRIPT instead of silently dropping the value; EOF is true when only blank lines remain; the FILE MODE ERROR wording is no longer inverted.	auto
9	Interactive INPUT of a number now parses like VAL (12ABC → 12, was 0). Real CoCo would print ?REDO and re-prompt; not implemented.	auto	Deduplicates the number heuristic.

Web server

#	Decision	Who	Notes
26	Server binds 127.0.0.1 by default; LAN exposure is opt-in via BASICOCO_HOST=0.0.0.0 (logs a no-authentication warning). Socket.IO accepts same-origin only (was *); extra origins via BASICOCO_CORS_ORIGINS. DEBUG=true with a non-loopback host refuses to start (Werkzeug debugger = remote code execution).	auto	Behaviour change: to reach the server from another machine you now must set BASICOCO_HOST.
27	Each interpreter has a command_lock; execute/input/continue/resume handlers run under it (@_exclusive), so a second command waits instead of racing. Ctrl+C never blocks: if a program is executing it sets break_requested, which the executor checks every statement and stops with BREAK IN n (CONT-able).	auto
27	At most 16 interpreters (tabs) per connection; the client now sends close_tab so closed tabs free their interpreter.	auto	Limit value is a judgment call.
28	Returning to a tab no longer auto-continues a program that is waiting for INPUT.	auto	Its live-server test was not re-run against the old code (it relies on the audit's probe of the bug).
30	A disconnected client's session is kept for 10 minutes (SESSION_GRACE_SECONDS); on reconnect the client offers its old session id (Socket.IO auth) and gets its tabs/programs back ("RECONNECTED - PROGRAMS KEPT"). Only orphaned sessions can be resumed; unknown ids get a fresh session. A program executing at disconnect is asked to stop.	auto	Session ids are random UUIDs.
29	Client: INKEY$ keys go to the active tab; keys are forwarded during any immediate command (not just RUN); a forwarded key isn't also typed into the command line; Ctrl+C resets the text display's input flag.	auto	JS changes are syntax-checked only — not exercised in a browser.
31	Client: SOUND notes are queued back-to-back on the AudioContext timeline (a melody in a loop plays every note); Ctrl+C silences queued notes.	auto	Same caveat: not exercised in a browser.
8	PCLS c clears to color c (server and client). DRAW M draws (BM doesn't). The server tracks drawn pixels (LINE, boxes, CIRCLE, DRAW, PSET/PRESET; clipped to 256×192) so PPOINT answers correctly.	auto	PAINT fills and GPRINT text are not pixel-tracked.
32	Log level from BASICOCO_LOG_LEVEL (default INFO); start_server_with_logging.sh defaults to DEBUG. Debug pretty-printing of output is lazy (no cost when DEBUG is off). Removed the ineffective FLASK_ENV/FLASK_DEBUG exports.	auto	Each start writes a new log file, so no rotation added; the 60 MB of old logs in logs/ can be pruned by hand.

Interpreter behaviour

#	Decision	Who	Notes
14	All program-line edits go through one method, CoCoBasic.store_program_line() (typed lines, DELETE, LOAD/MERGE, RENUM, app.py tab restore).	auto	Removes 3 duplicated copies.
16, 66	The runaway-loop guard (50k / 10M statements) budgets each uninterrupted slice of execution: the counter resets whenever a program resumes after INPUT, PAUSE or a graphics auto-yield.	auto	Consequence: an animation loop that yields every frame never trips the guard; Ctrl+C (task #27) is the way to stop it.
18, 43	NEXT honours its variable list (NEXT J,I) and unwinds abandoned inner loops (MS BASIC). Re-executing FOR for an active variable discards that loop — but only within the current GOSUB level, so a subroutine's LOCAL I + FOR I leaves the caller's FOR I intact (keeps the LOCAL extension's documented behaviour).	auto
21	RENUM refuses (error) when the new numbering would reorder lines, and also renumbers ELSE / RESTORE / RESUME targets.	auto
23	Malformed LOOP/IF conditions are syntax errors (were silently false). evaluate_condition() now propagates errors.	auto
24	RESUME accepts a label (labels win over same-named variables, as with GOTO).	auto

Parser / single-line control structures

#	Decision	Who	Notes
36, 37, 38, 42	Rewrote ast_converter.py as a text-level expander. Statements are split on colons and kept verbatim (no AST→text round trip, so parentheses can't be lost). Only one-line IFs are expanded (IF c THEN / body / [ELSE / body] / ENDIF); FOR/WHILE/DO lines are plain colon splits, since the executor already runs loops across sublines of one line.	auto	Fixes statements-after-NEXT, nested one-line FORs, dropped parens, crunched THEN, REM colons, ELSE binding.
42	ELSE binds to the nearest unmatched IF; THEN may be crunched ("X"THEN, 0THEN); IF c GOTO n, THEN n, ELSE n and THEN <expression> are implicit GOTOs.	auto	A branch is treated as a GOTO target when it doesn't start with a command word and has no top-level =/,/;.
36	Deliberate deviation from CoCo kept: inside a loop that opens and closes on the same line (FOR I=1 TO 3: IF I=2 THEN PRINT "TWO": NEXT I), an IF's body is only its own statement, so the NEXT isn't swallowed. Everywhere else an IF's body runs to the end of the line (CoCo).	auto — review	Existing tests and programs relied on this; real Color BASIC would put NEXT inside the IF.
—	A one-line IF whose branches are only GOTOs stays a single subline (no if_stack entry to leak on each taken jump, e.g. IF A$="" THEN 10).	auto	GOTO out of a multi-statement IF block still leaves a stale if_stack entry (pre-existing, documented).
40	Operator precedence now Color BASIC: OR < AND < NOT < relational (one level, left-to-right) < + − < MOD < * / < unary − < ^ (left-assoc). =<, =>, >< accepted.	user (dialect) + auto	NOT A=5 = NOT(A=5); -2^2 = −4; 2^3^2 = 64; 10 MOD 3*2 = 4.
40	MOD rounds operands to integers and truncates toward zero (-7 MOD 3 = −1).	auto	Microsoft BASIC semantics; CoCo itself has no MOD.
39	^ computed in floating point (so 9^9^9 → OVERFLOW instantly instead of hanging the server); integer results stay integers when exact. Overflow in + − * / ^ → OVERFLOW error; negative base to fractional power → ILLEGAL FUNCTION CALL. All these are BASIC errors (trappable by ON ERROR) via one shared BASIC_RUNTIME_ERRORS tuple in error_context.py.	auto
44	Numbers print CoCo-style: up to 9 significant digits, no leading 0 (.5), E notation outside .01 ≤ |x| < 1e9 (1.26765061E+30 style, trailing zeros dropped, 1E-03). Shared by PRINT, STR$, PRINT#: format_basic_number() in ast_nodes.py.	user (dialect) + auto	Exact CoCo rounding of the 9th digit may differ in edge cases.
41	Strict parsing: leftover tokens are a syntax error (X=5 6, PRINT 1 2); unknown characters are syntax errors (were silently dropped); &HFF/&O17 literals; ? = PRINT; ' starts a comment; 1.2.3 and 1E no longer leak Python messages.	auto	# is still tolerated by the tokenizer (file I/O is intercepted earlier).

Comments and crunched spacing (task #73, from the AST-usage survey #71)

#	Decision	Who	Notes
73	' is a comment (Extended Color BASIC shorthand for REM), both at the start of a line/statement and after a statement (SOUND 100,5 ' beep). Exception: after SAVE/LOAD/KILL/MERGE/CHAIN/CD a ' quotes a filename (SAVE 'GAME'), which existing tests and users rely on.	auto
73	REM matches by prefix everywhere (as the CoCo's keyword cruncher does): REMARK and REMEMBER=5 are comments. The splitter, is_rem_line and the tokenizer now agree (previously they disagreed).	auto	One old test asserting REMEMBER=5 splits as a variable was changed; another old test already asserted it's REM.
73	Comment sublines compile to a no-op instead of being re-dispatched as text on every execution.	auto
73	Block IF recognised when written IF(A=1)THEN or followed only by a comment (IF c THEN REM …, IF c THEN ' …).	auto
73	Accepted spacing variants: DATA"A", LINE INPUT, LOOP WHILE(X<3). LOOP followed by anything other than WHILE/UNTIL/comment is a syntax error.	auto
78	Crunched keywords generally (SOUND100,5, FORI=1TO10) are not supported — decided by the user: keep requiring spaces (see the #78 row below).	user

Immediate mode, CHAIN, RENUM and other parsing fixes

#	Decision	Who	Notes
19	An immediate-mode line with control structures now runs alongside the stored program as a temporary line −1 (ending in END), not in place of it. It is removed when execution finishes and is never LISTed or SAVEd. So IF X THEN GOTO 100 and INPUT inside an immediate loop now work.	auto	Replaces the save/restore-state approach, which discarded a pending INPUT.
19	GOTO n / GOSUB n typed at the prompt run the stored program from line n without clearing variables (CoCo behaviour). An immediate GOSUB's RETURN comes back to the prompt. An undefined target is UNDEFINED LINE. STOP typed at the prompt prints BREAK and does not arm CONT.	auto	Two old tests asserting a raw jump directive were updated.
20	CHAIN inside a running program returns a chain directive; the executor switches to the new program in the same loop (no nesting), so INPUT in a chained program resumes correctly and self-CHAIN loops don't hit Python's recursion limit. CHAIN typed at the prompt still starts a run.	auto
22	Position lookups use binary search (bisect). A 20k-iteration loop in a 5,000-line program went from 0.96 s to 0.03 s.	auto
24	Unknown command categories are created instead of dropped; RESUME moved to 'control', 'file' added — HELP lists every command.	auto
74	RENUM rewrites only real jump targets, via a quote/comment-aware scan (never text inside strings or REMs); handles crunched GOTO10.	auto
75	One graphics coordinate parser (_parse_coord_range) matches (x1,y1)-(x2,y2) by parenthesis depth; supports LINE -(x,y) from the last LINE end point; GET fixed; SCREEN/COLOR always split arguments properly; parenthesis matching ignores quoted text.	auto	CommandRegistry.is_coordinate_pair_syntax/parse_line_coordinates are now used only by their own tests (dead code, see #25).
76	File-number expressions may contain commas (PRINT #F(1,2),X); LINE INPUT A$(3) stores into the array element.	auto	PRINT# items are still parsed by file_io's own tokenizer, not the AST PRINT parser.
77	10PRINT accepted as a program line; DATA &HFF,&O17 read as numbers; ERR classified from the error's first line with quoted text removed (user DATA can't change ERR).	auto	Not true typed error codes yet.
—	continue_program_execution clears waiting_for_input itself (resuming means the answer arrived).	auto
47	BlockNode is live (multi-statement IF branches parse to it), so visit_block was fixed, not removed: it now stops at the first error, INPUT/PAUSE or flow-control directive (it checked a misspelled exit_for). ELSE <line> validates its target like THEN <line>. Removed as unused: ProgramNode, NodeType.PROGRAM, visit_program, StatementSplitter.expand_line_to_sublines (its 6 tests only re-tested split_on_delimiter cases that are already covered directly).	auto
25	Removed code with no production callers: core.clear_program, save_execution_state/restore_execution_state, StatementSplitter.has_control_keyword/CONTROL_KEYWORDS (superseded by ast_converter.starts_control_structure), CommandRegistry.parse_coordinates/is_coordinate_pair_syntax/parse_line_coordinates (superseded by graphics' _parse_coord_range), the unused pause_duration field. Their tests were ported to the replacements where the cases were not already covered (control-structure detection; 2-D array coordinates in LINE), not just deleted. PAUSE no longer does its own resume-point bookkeeping (the executor's pause handler already does it). run_program_from_line is now a thin wrapper over run_program(start_line=…), and every run resets the iteration counter.	auto
12	Graphics argument validation: COLOR outside 0-8, PMODE mode outside 0-4 or start page outside 1-8 → ILLEGAL FUNCTION CALL (with suggestions). SCREEN m,c: any nonzero color set means set 1, as on the CoCo (not an error). A string where a graphics command needs a number (PSET(A$,1)) → TYPE MISMATCH via eval_int, instead of a leaked Python int() message.	auto	Off-screen PSET/LINE coordinates are still clipped silently (#8), not an FC error. CIRCLE's ratio/arc arguments are still ignored; that feature is folded into #62.
12	INSTR(0,…) → ILLEGAL FUNCTION CALL (was silently clamped to 1; one old test updated). No upper limit on the start position, because this project's strings aren't capped at 255. LEN(number) → TYPE MISMATCH (was the length of its text).	auto	CoCo semantics per the dialect decision.
13	Error line attribution: immediate-mode errors show no line (was "at line 0", sometimes twice), as on the CoCo; program errors name their line once (parser errors used to say "at line 1"; PRINT-expression errors had no line). New ErrorContextManager.wrapped_error() is the one way to wrap an inner error: it keeps the inner first line and the inner suggestions, so no doubled "Suggestions:" block. Runtime errors in immediate statements are no longer labelled "SYNTAX ERROR" (X=Q(20) → BAD SUBSCRIPT).	auto
13	Every error site found with 0-1 suggestions now has at least 2, including EOF's errors (which used to be plain ValueErrors), file errors and graphics errors. Removed as unused: ErrorContextManager.push_context/pop_context/execution_stack/warning/get_stack_trace, the global error_context instance and the legacy dict helpers (create_legacy_error, syntax_error(), ... at module level), and VariableManager.set_variable/_calculate_linear_index/_get_default_value. The "unused import in functions.py" finding was stale: every import is used.	auto
45	Type checking (CoCo ?TM): strings may only be concatenated (+) or compared (= <> < > <= >=) with other strings; any other operator on a string, and any string/number mix, is TYPE MISMATCH (ERR 13, trappable by ON ERROR). "A"*2 used to give "AA" (Python semantics). Assignment checks the $ suffix: A = "X", A$ = 5, N(1) = "S" are TYPE MISMATCH (they used to be stored silently). FOR A$=…, string FOR bounds (even numeric-looking ones like "1", which used to be converted), and LINE INPUT into a numeric variable are TYPE MISMATCH too.	auto	Python-level string comparison (by code point) matches CoCo's ASCII ordering.
35	CLI: Ctrl+C used to disconnect the CLI from the server, because engine.io installs its own SIGINT handler; the client is now created with handle_sigint=False. Ctrl+C now sends BREAK during any command (not only RUN, so it works for GOTO 100 or an immediate loop), and the CLI stays connected. The 30 s RUN timeout (which returned to the prompt while the program was still printing) and the unbounded wait for other commands are replaced by one wait that ends on completion or on a dropped connection. CLS clears the terminal (the CLI listened for a clear message the server never sends). New e2e tests cover Ctrl+C and CLS.	auto
35	Web client (syntax-checked with node --check; not exercised in a browser): removed handlers for messages the server never sends (draw, clear_graphics, set_pmode) and the executeDraw stub; Copy copies the real text buffer (it used to copy a placeholder); the graphics info shows the real background color name (every non-black was "Green"); removed the SVG button, which only showed "coming soon", and the help lines advertising SVG export and "Share programs via URL".	auto	Left for review: saveSession/loadSession and the auto-save preference are half-built (loadSession is never called, and saved ImageData serializes to {}). Finish them or remove them? (Correction: the claim that the reconnect grace (#30) already covered a reload was wrong — the session id lived only in page memory, so a reload always started a fresh session.)
35	Session save/load replaced (user chose option 2, 2026-09-21): removed the half-built localStorage save (Save Session button, auto-save preference and timer, saveSession/loadSession). The session id and the tab strip (ids, titles, active tab, tab counter) now live in the browser tab's sessionStorage, so a reload within the 10-minute grace reconnects to the same server session: every tab's program, variables and state come back, and you land on the tab you were in. Restored tabs never push an empty state to the server, and new tabs never reuse an id the server still holds. Not covered: closing the browser tab, or reloading after 10 minutes (use SAVE).	user	Browser tests in tests/integration/browser/test_browser_reload.py (they failed before the change).
53	Tests strengthened: the pycuber gate (tools/validate_moves.py, 11 tests) is collected via tests/unit/test_validate_moves.py, so -m slow / -m "" run it; vacuous assertions (len(result) >= 0, bare isinstance(result, list), if result is not None:) replaced with real ones; the TIMER tests use a frozen clock instead of a ±10-tick tolerance; new tests for GET/PUT, GPRINT and CLEAR. Two bugs they found: CLEAR n printed "OK" in the middle of a running program (now silent while running), and PUT accepted any action word (now PSET/PRESET/AND/OR/NOT, else SYNTAX ERROR).	auto	The web client only draws PUT's PSET action; the others do nothing (added to #34).
34	GPRINT text is uppercased on the server (the 4×6 font, like the CoCo's, has no lowercase, so lowercase drew as "?"), and numbers use BASIC formatting (GPRINT(0,0),4/2 drew "2.0"). The other #34 items are client-only rendering fixes that need a browser to verify, so they were left for a supervised session.	auto
34	Client rendering fixed (user asked, 2026-09-21): CIRCLE is drawn with the server's midpoint algorithm in whole BASIC pixels (the antialiased arc left blended edges PAINT could leak through); PUT implements PRESET/AND/OR/NOT; a new tab gets proper default colors and tabs save/restore their colors, GET blocks and text history; the command being typed is kept in the scrollback buffer, wraps at 80 columns and leaves no stale cursor block; the GPRINT font gains # $ % & ; < = > @ [ \ ] ^ _.	user	The browser extension wasn't connected, so this was verified with tests/client/dual_monitor_harness.js: the real client code running in Node on a fake canvas with real pixels, cross-checked against the server's circle pixels (17 checks; the pre-fix client fails 16). Not yet looked at in a real browser.
69	rubiks_interactive.bas offers all nine turns (7=LEFT, 8=BOTTOM, 9=BACK added; the engine already had them), with a new test that drives the menu through INPUT and checks the exact turns. Left as they are: the unused S slice / X,Y,Z in the engine (they fold into #56's table-driven ApplyMove); rubiks_cube_rotate.bas and rubiks_test_moves.bas duplicating engine code, because they're deliberately self-contained demos that don't MERGE the library.	auto
78	Decided (user, 2026-09-21): keep requiring spaces around keywords. Crunched keywords (FORI=1TO3, GOTO10, SOUND100,5, IFX=1THENPRINT"Y") are SYNTAX ERRORs today (crunched line numbers like 10PRINT and "X"THEN do work, #77/#38). Real Color BASIC accepts them because its tokenizer matches keywords anywhere, and the price is that no variable name may contain a keyword (TOTAL holds TO, FORM holds FOR, SCORE holds OR). Adopting that would break names the bundled programs use. Recommendation: keep requiring spaces around keywords, and maybe add a clearer hint ("put a space after GOTO") when an unknown word starts with a keyword.	user	Recommendation accepted.
79	PMODE 0-4 follow the CoCo (user asked, 2026-09-21): every mode uses 0-255 × 0-191 coordinates; lower modes only have coarser pixels (PMODE 0/1: 2×2 coordinates per pixel, 2/3: 2×1). The client used to scale coordinates per mode, so in PMODE 1-3 most of the screen was drawn off-canvas. PMODE 0 is a graphics mode: it used to be treated as "text mode", so drawing after PMODE 0 gave ILLEGAL FUNCTION CALL. "No PMODE yet" is now None (server) / null (client). PPOINT snaps to the mode's pixel, as on the CoCo (in PMODE 1, PSET(1,1) lights (0,0) too). GPRINT draws each font pixel as one mode pixel, so text stays legible in coarse modes.	user	Checked by server tests, the Node harness (16 new checks) and headless-Chrome tests in PMODE 0/1/3/4.
25	min_args / max_args on CommandRegistry.register are left in place but not enforced: no command passes them, and enforcing them would mean auditing every handler's argument count.	auto
25	Dropped (user, 2026-09-21): min_args / max_args removed from CommandRegistry.register. Each handler already validates its own arguments with command-specific errors, and a generic comma count fits BASIC syntax badly (LINE(0,0)-(9,9),PSET,BF).	user

Rubik's cube programs

#	Decision	Who	Notes
63	Step 6 no longer applies a net U rotation; RotateCorrectToTFR's setup turn is undone after the algorithm (UD$).	auto	Audit's verified patch; 1,519/1,519 scrambles solve.
64	Step 1 flipped-edge insertions: TI=1 bR, TI=3 fLF.	auto	Found by brute-force search against pycuber.
65	Scramble uses face turns only (R U F L D B).	auto
51	AlignMidEdge scratch variable renamed SC → SK (SC is the render scale).	auto
67	R and L animation rotation signs swapped. ROT-X keeps its formula (it turns like L).	auto	The ISSUES.md z-sort diagnosis was wrong; to be updated with the docs task.

Tests / tooling

#	Decision	Who	Notes
48	testpaths = tests (was tests/unit).	auto
49	8 old e2e/cli files removed (git rm); replaced by tests/integration/e2e/test_cli_sessions.py + a live_server fixture that starts app.py in a temp dir on a free port.	user approved removal
50, 33	pycuber, pytest-timeout added to requirements.txt.	auto
33	Werkzeug==3.1.6 and simple-websocket==1.1.0 pinned at the versions installed and tested (Werkzeug was unpinned; Flask 2.3 accepts anything ≥ 2.3.7). The Flask 2.3 / Flask-SocketIO 5.3 stack was not upgraded: that needs package installs, so it waits for approval.	auto
33	Upgraded (user, 2026-09-21) to Flask 3.1.3, Flask-SocketIO 5.6.1, python-socketio 5.17.0, python-engineio 4.14.0, Werkzeug 3.1.8. No code changes were needed; the full suite (1948 tests, including live-server, websocket and CLI e2e) passes.	user	Previous environment saved as a pip freeze in the session scratchpad.
—	tests/unit/test_utilities.py removed; stray programs/CHARTDATA.DAT removed (backed up to the session scratchpad).	user